Introduction to Data Mining presents fundamental concepts and algorithms for those learning data mining for the first ti
1,147 196 13MB
English Pages 736 [737] Year 2013
Table of contents :
Cover......Page 1
Table of Contents......Page 4
Chapter 1. Introduction......Page 6
Chapter 2. Data......Page 24
Chapter 3. Exploring Data......Page 102
Chapter 4. Classification: Basic Concepts, Decision Trees, and Model Evaluation......Page 150
Chapter 5. Classification: Alternative Techniques......Page 212
Chapter 6. Association Analysis: Basic Concepts and Algorithms......Page 332
Chapter 7. Association Analysis: Advanced Concepts......Page 420
Chapter 8. Cluster Analysis: Basic Concepts and Algorithms......Page 492
Chapter 9. Cluster Analysis: Additional Issues and Algorithms......Page 574
Chapter 10. Anomaly Detection......Page 656
Appendix B: Dimensionality Reduction......Page 690
Appendix D: Regression......Page 708
Appendix E: Optimization......Page 718
Copyright Permissions......Page 729
A......Page 730
C......Page 731
E......Page 732
M......Page 733
P......Page 734
S......Page 735
W......Page 736
Z......Page 737
Introduction to Data Mining Tan et al. First Edition
ISBN 9781292026152
9 781292 026152
Introduction to Data Mining Tan Steinbach Kumar First Edition
Introduction to Data Mining Tan Steinbach Kumar First Edition
Pearson Education Limited Edinburgh Gate Harlow Essex CM20 2JE England and Associated Companies throughout the world Visit us on the World Wide Web at: www.pearsoned.co.uk © Pearson Education Limited 2014 All rights reserved. No part of this publication may be reproduced, stored in a retrieval system, or transmitted in any form or by any means, electronic, mechanical, photocopying, recording or otherwise, without either the prior written permission of the publisher or a licence permitting restricted copying in the United Kingdom issued by the Copyright Licensing Agency Ltd, Saffron House, 6–10 Kirby Street, London EC1N 8TS. All trademarks used herein are the property of their respective owners. The use of any trademark in this text does not vest in the author or publisher any trademark ownership rights in such trademarks, nor does the use of such trademarks imply any afﬁliation with or endorsement of this book by such owners.
ISBN 10: 1292026154 ISBN 13: 9781292026152
British Library CataloguinginPublication Data A catalogue record for this book is available from the British Library Printed in the United States of America
P
E
A
R
S
O
N
C U
S T O
M
L
I
B
R
A
R Y
Table of Contents Chapter 1. Introduction PangNing Tan/Michael Steinbach/Vipin Kumar
1
Chapter 2. Data PangNing Tan/Michael Steinbach/Vipin Kumar
19
Chapter 3. Exploring Data PangNing Tan/Michael Steinbach/Vipin Kumar
97
Chapter 4. Classification: Basic Concepts, Decision Trees, and Model Evaluation PangNing Tan/Michael Steinbach/Vipin Kumar
145
Chapter 5. Classification: Alternative Techniques PangNing Tan/Michael Steinbach/Vipin Kumar
207
Chapter 6. Association Analysis: Basic Concepts and Algorithms PangNing Tan/Michael Steinbach/Vipin Kumar
327
Chapter 7. Association Analysis: Advanced Concepts PangNing Tan/Michael Steinbach/Vipin Kumar
415
Chapter 8. Cluster Analysis: Basic Concepts and Algorithms PangNing Tan/Michael Steinbach/Vipin Kumar
487
Chapter 9. Cluster Analysis: Additional Issues and Algorithms PangNing Tan/Michael Steinbach/Vipin Kumar
569
Chapter 10. Anomaly Detection PangNing Tan/Michael Steinbach/Vipin Kumar
651
Appendix B: Dimensionality Reduction PangNing Tan/Michael Steinbach/Vipin Kumar
685
Appendix D: Regression PangNing Tan/Michael Steinbach/Vipin Kumar
703
Appendix E: Optimization PangNing Tan/Michael Steinbach/Vipin Kumar
713
I
II
Copyright Permissions PangNing Tan/Michael Steinbach/Vipin Kumar
724
Index
725
1 Introduction Rapid advances in data collection and storage technology have enabled organizations to accumulate vast amounts of data. However, extracting useful information has proven extremely challenging. Often, traditional data analysis tools and techniques cannot be used because of the massive size of a data set. Sometimes, the nontraditional nature of the data means that traditional approaches cannot be applied even if the data set is relatively small. In other situations, the questions that need to be answered cannot be addressed using existing data analysis techniques, and thus, new methods need to be developed. Data mining is a technology that blends traditional data analysis methods with sophisticated algorithms for processing large volumes of data. It has also opened up exciting opportunities for exploring and analyzing new types of data and for analyzing old types of data in new ways. In this introductory chapter, we present an overview of data mining and outline the key topics to be covered in this book. We start with a description of some wellknown applications that require new techniques for data analysis. Business Pointofsale data collection (bar code scanners, radio frequency identiﬁcation (RFID), and smart card technology) have allowed retailers to collect uptotheminute data about customer purchases at the checkout counters of their stores. Retailers can utilize this information, along with other businesscritical data such as Web logs from ecommerce Web sites and customer service records from call centers, to help them better understand the needs of their customers and make more informed business decisions. Data mining techniques can be used to support a wide range of business intelligence applications such as customer proﬁling, targeted marketing, workﬂow management, store layout, and fraud detection. It can also help retailers
From Chapter 1 of Introduction to Data Mining, First Edition. PangNing Tan, Michael Steinbach, Vipin Kumar. Copyright © 2006 by Pearson Education, Inc. All rights reserved.
1
Chapter 1
Introduction
answer important business questions such as “Who are the most proﬁtable customers?” “What products can be crosssold or upsold?” and “What is the revenue outlook of the company for next year?” Some of these questions motivated the creation of association analysis (Chapters 6 and 7), a new data analysis technique. Medicine, Science, and Engineering Researchers in medicine, science, and engineering are rapidly accumulating data that is key to important new discoveries. For example, as an important step toward improving our understanding of the Earth’s climate system, NASA has deployed a series of Earthorbiting satellites that continuously generate global observations of the land surface, oceans, and atmosphere. However, because of the size and spatiotemporal nature of the data, traditional methods are often not suitable for analyzing these data sets. Techniques developed in data mining can aid Earth scientists in answering questions such as “What is the relationship between the frequency and intensity of ecosystem disturbances such as droughts and hurricanes to global warming?” “How is land surface precipitation and temperature aﬀected by ocean surface temperature?” and “How well can we predict the beginning and end of the growing season for a region?” As another example, researchers in molecular biology hope to use the large amounts of genomic data currently being gathered to better understand the structure and function of genes. In the past, traditional methods in molecular biology allowed scientists to study only a few genes at a time in a given experiment. Recent breakthroughs in microarray technology have enabled scientists to compare the behavior of thousands of genes under various situations. Such comparisons can help determine the function of each gene and perhaps isolate the genes responsible for certain diseases. However, the noisy and highdimensional nature of data requires new types of data analysis. In addition to analyzing gene array data, data mining can also be used to address other important biological challenges such as protein structure prediction, multiple sequence alignment, the modeling of biochemical pathways, and phylogenetics.
1.1
What Is Data Mining?
Data mining is the process of automatically discovering useful information in large data repositories. Data mining techniques are deployed to scour large databases in order to ﬁnd novel and useful patterns that might otherwise remain unknown. They also provide capabilities to predict the outcome of a
2
1.1
What Is Data Mining?
future observation, such as predicting whether a newly arrived customer will spend more than $100 at a department store. Not all information discovery tasks are considered to be data mining. For example, looking up individual records using a database management system or ﬁnding particular Web pages via a query to an Internet search engine are tasks related to the area of information retrieval. Although such tasks are important and may involve the use of the sophisticated algorithms and data structures, they rely on traditional computer science techniques and obvious features of the data to create index structures for eﬃciently organizing and retrieving information. Nonetheless, data mining techniques have been used to enhance information retrieval systems. Data Mining and Knowledge Discovery Data mining is an integral part of knowledge discovery in databases (KDD), which is the overall process of converting raw data into useful information, as shown in Figure 1.1. This process consists of a series of transformation steps, from data preprocessing to postprocessing of data mining results. Input Data
Data Preprocessing
Data Mining
Feature Selection Dimensionality Reduction Normalization Data Subsetting
Postprocessing
Information
Filtering Patterns Visualization Pattern Interpretation
Figure 1.1. The process of knowledge discovery in databases (KDD).
The input data can be stored in a variety of formats (ﬂat ﬁles, spreadsheets, or relational tables) and may reside in a centralized data repository or be distributed across multiple sites. The purpose of preprocessing is to transform the raw input data into an appropriate format for subsequent analysis. The steps involved in data preprocessing include fusing data from multiple sources, cleaning data to remove noise and duplicate observations, and selecting records and features that are relevant to the data mining task at hand. Because of the many ways data can be collected and stored, data
3
Chapter 1
Introduction
preprocessing is perhaps the most laborious and timeconsuming step in the overall knowledge discovery process. “Closing the loop” is the phrase often used to refer to the process of integrating data mining results into decision support systems. For example, in business applications, the insights oﬀered by data mining results can be integrated with campaign management tools so that eﬀective marketing promotions can be conducted and tested. Such integration requires a postprocessing step that ensures that only valid and useful results are incorporated into the decision support system. An example of postprocessing is visualization (see Chapter 3), which allows analysts to explore the data and the data mining results from a variety of viewpoints. Statistical measures or hypothesis testing methods can also be applied during postprocessing to eliminate spurious data mining results.
1.2
Motivating Challenges
As mentioned earlier, traditional data analysis techniques have often encountered practical diﬃculties in meeting the challenges posed by new data sets. The following are some of the speciﬁc challenges that motivated the development of data mining. Scalability Because of advances in data generation and collection, data sets with sizes of gigabytes, terabytes, or even petabytes are becoming common. If data mining algorithms are to handle these massive data sets, then they must be scalable. Many data mining algorithms employ special search strategies to handle exponential search problems. Scalability may also require the implementation of novel data structures to access individual records in an efﬁcient manner. For instance, outofcore algorithms may be necessary when processing data sets that cannot ﬁt into main memory. Scalability can also be improved by using sampling or developing parallel and distributed algorithms. High Dimensionality It is now common to encounter data sets with hundreds or thousands of attributes instead of the handful common a few decades ago. In bioinformatics, progress in microarray technology has produced gene expression data involving thousands of features. Data sets with temporal or spatial components also tend to have high dimensionality. For example, consider a data set that contains measurements of temperature at various locations. If the temperature measurements are taken repeatedly for an extended period, the number of dimensions (features) increases in proportion to
4
1.2
Motivating Challenges
the number of measurements taken. Traditional data analysis techniques that were developed for lowdimensional data often do not work well for such highdimensional data. Also, for some data analysis algorithms, the computational complexity increases rapidly as the dimensionality (the number of features) increases. Heterogeneous and Complex Data Traditional data analysis methods often deal with data sets containing attributes of the same type, either continuous or categorical. As the role of data mining in business, science, medicine, and other ﬁelds has grown, so has the need for techniques that can handle heterogeneous attributes. Recent years have also seen the emergence of more complex data objects. Examples of such nontraditional types of data include collections of Web pages containing semistructured text and hyperlinks; DNA data with sequential and threedimensional structure; and climate data that consists of time series measurements (temperature, pressure, etc.) at various locations on the Earth’s surface. Techniques developed for mining such complex objects should take into consideration relationships in the data, such as temporal and spatial autocorrelation, graph connectivity, and parentchild relationships between the elements in semistructured text and XML documents. Data Ownership and Distribution Sometimes, the data needed for an analysis is not stored in one location or owned by one organization. Instead, the data is geographically distributed among resources belonging to multiple entities. This requires the development of distributed data mining techniques. Among the key challenges faced by distributed data mining algorithms include (1) how to reduce the amount of communication needed to perform the distributed computation, (2) how to eﬀectively consolidate the data mining results obtained from multiple sources, and (3) how to address data security issues. Nontraditional Analysis The traditional statistical approach is based on a hypothesizeandtest paradigm. In other words, a hypothesis is proposed, an experiment is designed to gather the data, and then the data is analyzed with respect to the hypothesis. Unfortunately, this process is extremely laborintensive. Current data analysis tasks often require the generation and evaluation of thousands of hypotheses, and consequently, the development of some data mining techniques has been motivated by the desire to automate the process of hypothesis generation and evaluation. Furthermore, the data sets analyzed in data mining are typically not the result of a carefully designed
5
Chapter 1
Introduction
experiment and often represent opportunistic samples of the data, rather than random samples. Also, the data sets frequently involve nontraditional types of data and data distributions.
1.3
The Origins of Data Mining
Brought together by the goal of meeting the challenges of the previous section, researchers from diﬀerent disciplines began to focus on developing more eﬃcient and scalable tools that could handle diverse types of data. This work, which culminated in the ﬁeld of data mining, built upon the methodology and algorithms that researchers had previously used. In particular, data mining draws upon ideas, such as (1) sampling, estimation, and hypothesis testing from statistics and (2) search algorithms, modeling techniques, and learning theories from artiﬁcial intelligence, pattern recognition, and machine learning. Data mining has also been quick to adopt ideas from other areas, including optimization, evolutionary computing, information theory, signal processing, visualization, and information retrieval. A number of other areas also play key supporting roles. In particular, database systems are needed to provide support for eﬃcient storage, indexing, and query processing. Techniques from high performance (parallel) computing are often important in addressing the massive size of some data sets. Distributed techniques can also help address the issue of size and are essential when the data cannot be gathered in one location. Figure 1.2 shows the relationship of data mining to other areas.
Statistics Data Mining
AI, Machine Learning, and Pattern Recognition
Database Technology, Parallel Computing, Distributed Computing
Figure 1.2. Data mining as a confluence of many disciplines.
6
1.4
1.4
Data Mining Tasks
Data Mining Tasks
Data mining tasks are generally divided into two major categories: Predictive tasks. The objective of these tasks is to predict the value of a particular attribute based on the values of other attributes. The attribute to be predicted is commonly known as the target or dependent variable, while the attributes used for making the prediction are known as the explanatory or independent variables. Descriptive tasks. Here, the objective is to derive patterns (correlations, trends, clusters, trajectories, and anomalies) that summarize the underlying relationships in data. Descriptive data mining tasks are often exploratory in nature and frequently require postprocessing techniques to validate and explain the results. Figure 1.3 illustrates four of the core data mining tasks that are described in the remainder of this book.
Data C An lust aly er sis
ID
Home Owner
Marital Status
Annual Income
Defaulted Borrower
1
Yes
Single
125K
No
2
No
Married
100K
No
3
No
Single
70K
No
4
Yes
Married
120K
No
5
No
Divorced 95K
Yes
6
No
Married
No
7
Yes
Divorced 220K
No
8
No
Single
85K
Yes
9
No
Married
75K
No
10
No
Single
90K
Yes
ion iat oc lysis s As na A
80K
ive ict g ed elin r P od M
A De nom tec aly tio n
DIAPER
DIAPER
Figure 1.3. Four of the core data mining tasks.
7
Chapter 1
Introduction
Predictive modeling refers to the task of building a model for the target variable as a function of the explanatory variables. There are two types of predictive modeling tasks: classiﬁcation, which is used for discrete target variables, and regression, which is used for continuous target variables. For example, predicting whether a Web user will make a purchase at an online bookstore is a classiﬁcation task because the target variable is binaryvalued. On the other hand, forecasting the future price of a stock is a regression task because price is a continuousvalued attribute. The goal of both tasks is to learn a model that minimizes the error between the predicted and true values of the target variable. Predictive modeling can be used to identify customers that will respond to a marketing campaign, predict disturbances in the Earth’s ecosystem, or judge whether a patient has a particular disease based on the results of medical tests. Example 1.1 (Predicting the Type of a Flower). Consider the task of predicting a species of ﬂower based on the characteristics of the ﬂower. In particular, consider classifying an Iris ﬂower as to whether it belongs to one of the following three Iris species: Setosa, Versicolour, or Virginica. To perform this task, we need a data set containing the characteristics of various ﬂowers of these three species. A data set with this type of information is the wellknown Iris data set from the UCI Machine Learning Repository at http://www.ics.uci.edu/∼mlearn. In addition to the species of a ﬂower, this data set contains four other attributes: sepal width, sepal length, petal length, and petal width. (The Iris data set and its attributes are described further in Section 3.1.) Figure 1.4 shows a plot of petal width versus petal length for the 150 ﬂowers in the Iris data set. Petal width is broken into the categories low, medium, and high, which correspond to the intervals [0, 0.75), [0.75, 1.75), [1.75, ∞), respectively. Also, petal length is broken into categories low, medium, and high, which correspond to the intervals [0, 2.5), [2.5, 5), [5, ∞), respectively. Based on these categories of petal width and length, the following rules can be derived: Petal width low and petal length low implies Setosa. Petal width medium and petal length medium implies Versicolour. Petal width high and petal length high implies Virginica.
While these rules do not classify all the ﬂowers, they do a good (but not perfect) job of classifying most of the ﬂowers. Note that ﬂowers from the Setosa species are well separated from the Versicolour and Virginica species with respect to petal width and length, but the latter two species overlap somewhat with respect to these attributes.
8
1.4
Data Mining Tasks
2.5 Setosa Versicolour Virginica
Petal Width (cm)
2 1.75 1.5
1 0.75 0.5
0 0
1
2
2.5
3
4
5
6
7
Petal Length (cm)
Figure 1.4. Petal width versus petal length for 150 Iris flowers.
Association analysis is used to discover patterns that describe strongly associated features in the data. The discovered patterns are typically represented in the form of implication rules or feature subsets. Because of the exponential size of its search space, the goal of association analysis is to extract the most interesting patterns in an eﬃcient manner. Useful applications of association analysis include ﬁnding groups of genes that have related functionality, identifying Web pages that are accessed together, or understanding the relationships between diﬀerent elements of Earth’s climate system. Example 1.2 (Market Basket Analysis). The transactions shown in Table 1.1 illustrate pointofsale data collected at the checkout counters of a grocery store. Association analysis can be applied to ﬁnd items that are frequently bought together by customers. For example, we may discover the rule {Diapers} −→ {Milk}, which suggests that customers who buy diapers also tend to buy milk. This type of rule can be used to identify potential crossselling opportunities among related items. Cluster analysis seeks to ﬁnd groups of closely related observations so that observations that belong to the same cluster are more similar to each other
9
Chapter 1
Introduction
Transaction ID 1 2 3 4 5 6 7 8 9 10
Table 1.1. Market basket data. Items {Bread, Butter, Diapers, Milk} {Coffee, Sugar, Cookies, Salmon} {Bread, Butter, Coffee, Diapers, Milk, Eggs} {Bread, Butter, Salmon, Chicken} {Eggs, Bread, Butter} {Salmon, Diapers, Milk} {Bread, Tea, Sugar, Eggs} {Coffee, Sugar, Chicken, Eggs} {Bread, Diapers, Milk, Salt} {Tea, Eggs, Cookies, Diapers, Milk}
than observations that belong to other clusters. Clustering has been used to group sets of related customers, ﬁnd areas of the ocean that have a signiﬁcant impact on the Earth’s climate, and compress data. Example 1.3 (Document Clustering). The collection of news articles shown in Table 1.2 can be grouped based on their respective topics. Each article is represented as a set of wordfrequency pairs (w, c), where w is a word and c is the number of times the word appears in the article. There are two natural clusters in the data set. The ﬁrst cluster consists of the ﬁrst four articles, which correspond to news about the economy, while the second cluster contains the last four articles, which correspond to news about health care. A good clustering algorithm should be able to identify these two clusters based on the similarity between words that appear in the articles. Table 1.2. Collection of news articles. Article 1 2 3 4 5 6 7 8
10
Words dollar: 1, industry: 4, country: 2, loan: 3, deal: 2, government: 2 machinery: 2, labor: 3, market: 4, industry: 2, work: 3, country: 1 job: 5, inﬂation: 3, rise: 2, jobless: 2, market: 3, country: 2, index: 3 domestic: 3, forecast: 2, gain: 1, market: 2, sale: 3, price: 2 patient: 4, symptom: 2, drug: 3, health: 2, clinic: 2, doctor: 2 pharmaceutical: 2, company: 3, drug: 2, vaccine: 1, ﬂu: 3 death: 2, cancer: 4, drug: 3, public: 4, health: 3, director: 2 medical: 2, cost: 3, increase: 2, patient: 2, health: 3, care: 1
1.5
Scope and Organization of the Book
Anomaly detection is the task of identifying observations whose characteristics are signiﬁcantly diﬀerent from the rest of the data. Such observations are known as anomalies or outliers. The goal of an anomaly detection algorithm is to discover the real anomalies and avoid falsely labeling normal objects as anomalous. In other words, a good anomaly detector must have a high detection rate and a low false alarm rate. Applications of anomaly detection include the detection of fraud, network intrusions, unusual patterns of disease, and ecosystem disturbances. Example 1.4 (Credit Card Fraud Detection). A credit card company records the transactions made by every credit card holder, along with personal information such as credit limit, age, annual income, and address. Since the number of fraudulent cases is relatively small compared to the number of legitimate transactions, anomaly detection techniques can be applied to build a proﬁle of legitimate transactions for the users. When a new transaction arrives, it is compared against the proﬁle of the user. If the characteristics of the transaction are very diﬀerent from the previously created proﬁle, then the transaction is ﬂagged as potentially fraudulent.
1.5
Scope and Organization of the Book
This book introduces the major principles and techniques used in data mining from an algorithmic perspective. A study of these principles and techniques is essential for developing a better understanding of how data mining technology can be applied to various kinds of data. This book also serves as a starting point for readers who are interested in doing research in this ﬁeld. We begin the technical discussion of this book with a chapter on data (Chapter 2), which discusses the basic types of data, data quality, preprocessing techniques, and measures of similarity and dissimilarity. Although this material can be covered quickly, it provides an essential foundation for data analysis. Chapter 3, on data exploration, discusses summary statistics, visualization techniques, and OnLine Analytical Processing (OLAP). These techniques provide the means for quickly gaining insight into a data set. Chapters 4 and 5 cover classiﬁcation. Chapter 4 provides a foundation by discussing decision tree classiﬁers and several issues that are important to all classiﬁcation: overﬁtting, performance evaluation, and the comparison of diﬀerent classiﬁcation models. Using this foundation, Chapter 5 describes a number of other important classiﬁcation techniques: rulebased systems, nearestneighbor classiﬁers, Bayesian classiﬁers, artiﬁcial neural networks, support vector machines, and ensemble classiﬁers, which are collections of classi
11
Chapter 1
Introduction
ﬁers. The multiclass and imbalanced class problems are also discussed. These topics can be covered independently. Association analysis is explored in Chapters 6 and 7. Chapter 6 describes the basics of association analysis: frequent itemsets, association rules, and some of the algorithms used to generate them. Speciﬁc types of frequent itemsets—maximal, closed, and hyperclique—that are important for data mining are also discussed, and the chapter concludes with a discussion of evaluation measures for association analysis. Chapter 7 considers a variety of more advanced topics, including how association analysis can be applied to categorical and continuous data or to data that has a concept hierarchy. (A concept hierarchy is a hierarchical categorization of objects, e.g., store items, clothing, shoes, sneakers.) This chapter also describes how association analysis can be extended to ﬁnd sequential patterns (patterns involving order), patterns in graphs, and negative relationships (if one item is present, then the other is not). Cluster analysis is discussed in Chapters 8 and 9. Chapter 8 ﬁrst describes the diﬀerent types of clusters and then presents three speciﬁc clustering techniques: Kmeans, agglomerative hierarchical clustering, and DBSCAN. This is followed by a discussion of techniques for validating the results of a clustering algorithm. Additional clustering concepts and techniques are explored in Chapter 9, including fuzzy and probabilistic clustering, SelfOrganizing Maps (SOM), graphbased clustering, and densitybased clustering. There is also a discussion of scalability issues and factors to consider when selecting a clustering algorithm. The last chapter, Chapter 10, is on anomaly detection. After some basic deﬁnitions, several diﬀerent types of anomaly detection are considered: statistical, distancebased, densitybased, and clusteringbased. Appendices A through E give a brief review of important topics that are used in portions of the book: linear algebra, dimensionality reduction, statistics, regression, and optimization. The subject of data mining, while relatively young compared to statistics or machine learning, is already too large to cover in a single book. Selected references to topics that are only brieﬂy covered, such as data quality, are provided in the bibliographic notes of the appropriate chapter. References to topics not covered in this book, such as data mining for streams and privacypreserving data mining, are provided in the bibliographic notes of this chapter.
12
1.6
1.6
Bibliographic Notes
Bibliographic Notes
The topic of data mining has inspired many textbooks. Introductory textbooks include those by Dunham [10], Han and Kamber [21], Hand et al. [23], and Roiger and Geatz [36]. Data mining books with a stronger emphasis on business applications include the works by Berry and Linoﬀ [2], Pyle [34], and Parr Rud [33]. Books with an emphasis on statistical learning include those by Cherkassky and Mulier [6], and Hastie et al. [24]. Some books with an emphasis on machine learning or pattern recognition are those by Duda et al. [9], Kantardzic [25], Mitchell [31], Webb [41], and Witten and Frank [42]. There are also some more specialized books: Chakrabarti [4] (web mining), Fayyad et al. [13] (collection of early articles on data mining), Fayyad et al. [11] (visualization), Grossman et al. [18] (science and engineering), Kargupta and Chan [26] (distributed data mining), Wang et al. [40] (bioinformatics), and Zaki and Ho [44] (parallel data mining). There are several conferences related to data mining. Some of the main conferences dedicated to this ﬁeld include the ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD), the IEEE International Conference on Data Mining (ICDM), the SIAM International Conference on Data Mining (SDM), the European Conference on Principles and Practice of Knowledge Discovery in Databases (PKDD), and the PaciﬁcAsia Conference on Knowledge Discovery and Data Mining (PAKDD). Data mining papers can also be found in other major conferences such as the ACM SIGMOD/PODS conference, the International Conference on Very Large Data Bases (VLDB), the Conference on Information and Knowledge Management (CIKM), the International Conference on Data Engineering (ICDE), the International Conference on Machine Learning (ICML), and the National Conference on Artiﬁcial Intelligence (AAAI). Journal publications on data mining include IEEE Transactions on Knowledge and Data Engineering, Data Mining and Knowledge Discovery, Knowledge and Information Systems, Intelligent Data Analysis, Information Systems, and the Journal of Intelligent Information Systems. There have been a number of general articles on data mining that deﬁne the ﬁeld or its relationship to other ﬁelds, particularly statistics. Fayyad et al. [12] describe data mining and how it ﬁts into the total knowledge discovery process. Chen et al. [5] give a database perspective on data mining. Ramakrishnan and Grama [35] provide a general discussion of data mining and present several viewpoints. Hand [22] describes how data mining diﬀers from statistics, as does Friedman [14]. Lambert [29] explores the use of statistics for large data sets and provides some comments on the respective roles of data mining and statistics.
13
Chapter 1
Introduction
Glymour et al. [16] consider the lessons that statistics may have for data mining. Smyth et al. [38] describe how the evolution of data mining is being driven by new types of data and applications, such as those involving streams, graphs, and text. Emerging applications in data mining are considered by Han et al. [20] and Smyth [37] describes some research challenges in data mining. A discussion of how developments in data mining research can be turned into practical tools is given by Wu et al. [43]. Data mining standards are the subject of a paper by Grossman et al. [17]. Bradley [3] discusses how data mining algorithms can be scaled to large data sets. With the emergence of new data mining applications have come new challenges that need to be addressed. For instance, concerns about privacy breaches as a result of data mining have escalated in recent years, particularly in application domains such as Web commerce and health care. As a result, there is growing interest in developing data mining algorithms that maintain user privacy. Developing techniques for mining encrypted or randomized data is known as privacypreserving data mining. Some general references in this area include papers by Agrawal and Srikant [1], Clifton et al. [7] and Kargupta et al. [27]. Vassilios et al. [39] provide a survey. Recent years have witnessed a growing number of applications that rapidly generate continuous streams of data. Examples of stream data include network traﬃc, multimedia streams, and stock prices. Several issues must be considered when mining data streams, such as the limited amount of memory available, the need for online analysis, and the change of the data over time. Data mining for stream data has become an important area in data mining. Some selected publications are Domingos and Hulten [8] (classiﬁcation), Giannella et al. [15] (association analysis), Guha et al. [19] (clustering), Kifer et al. [28] (change detection), Papadimitriou et al. [32] (time series), and Law et al. [30] (dimensionality reduction).
Bibliography [1] R. Agrawal and R. Srikant. Privacypreserving data mining. In Proc. of 2000 ACMSIGMOD Intl. Conf. on Management of Data, pages 439–450, Dallas, Texas, 2000. ACM Press. [2] M. J. A. Berry and G. Linoﬀ. Data Mining Techniques: For Marketing, Sales, and Customer Relationship Management. Wiley Computer Publishing, 2nd edition, 2004. [3] P. S. Bradley, J. Gehrke, R. Ramakrishnan, and R. Srikant. Scaling mining algorithms to large databases. Communications of the ACM, 45(8):38–43, 2002. [4] S. Chakrabarti. Mining the Web: Discovering Knowledge from Hypertext Data. Morgan Kaufmann, San Francisco, CA, 2003.
14
Bibliography [5] M.S. Chen, J. Han, and P. S. Yu. Data Mining: An Overview from a Database Perspective. IEEE Transactions on Knowledge abd Data Engineering, 8(6):866–883, 1996. [6] V. Cherkassky and F. Mulier. Learning from Data: Concepts, Theory, and Methods. Wiley Interscience, 1998. [7] C. Clifton, M. Kantarcioglu, and J. Vaidya. Deﬁning privacy for data mining. In National Science Foundation Workshop on Next Generation Data Mining, pages 126– 133, Baltimore, MD, November 2002. [8] P. Domingos and G. Hulten. Mining highspeed data streams. In Proc. of the 6th Intl. Conf. on Knowledge Discovery and Data Mining, pages 71–80, Boston, Massachusetts, 2000. ACM Press. [9] R. O. Duda, P. E. Hart, and D. G. Stork. Pattern Classification. John Wiley & Sons, Inc., New York, 2nd edition, 2001. [10] M. H. Dunham. Data Mining: Introductory and Advanced Topics. Prentice Hall, 2002. [11] U. M. Fayyad, G. G. Grinstein, and A. Wierse, editors. Information Visualization in Data Mining and Knowledge Discovery. Morgan Kaufmann Publishers, San Francisco, CA, September 2001. [12] U. M. Fayyad, G. PiatetskyShapiro, and P. Smyth. From Data Mining to Knowledge Discovery: An Overview. In Advances in Knowledge Discovery and Data Mining, pages 1–34. AAAI Press, 1996. [13] U. M. Fayyad, G. PiatetskyShapiro, P. Smyth, and R. Uthurusamy, editors. Advances in Knowledge Discovery and Data Mining. AAAI/MIT Press, 1996. [14] J. H. Friedman. Data Mining and Statistics: What’s the Connection? Unpublished. wwwstat.stanford.edu/∼jhf/ftp/dmstat.ps, 1997. [15] C. Giannella, J. Han, J. Pei, X. Yan, and P. S. Yu. Mining Frequent Patterns in Data Streams at Multiple Time Granularities. In H. Kargupta, A. Joshi, K. Sivakumar, and Y. Yesha, editors, Next Generation Data Mining, pages 191–212. AAAI/MIT, 2003. [16] C. Glymour, D. Madigan, D. Pregibon, and P. Smyth. Statistical Themes and Lessons for Data Mining. Data Mining and Knowledge Discovery, 1(1):11–28, 1997. [17] R. L. Grossman, M. F. Hornick, and G. Meyer. Data mining standards initiatives. Communications of the ACM, 45(8):59–61, 2002. [18] R. L. Grossman, C. Kamath, P. Kegelmeyer, V. Kumar, and R. Namburu, editors. Data Mining for Scientific and Engineering Applications. Kluwer Academic Publishers, 2001. [19] S. Guha, A. Meyerson, N. Mishra, R. Motwani, and L. O’Callaghan. Clustering Data Streams: Theory and Practice. IEEE Transactions on Knowledge and Data Engineering, 15(3):515–528, May/June 2003. [20] J. Han, R. B. Altman, V. Kumar, H. Mannila, and D. Pregibon. Emerging scientiﬁc applications in data mining. Communications of the ACM, 45(8):54–58, 2002. [21] J. Han and M. Kamber. Data Mining: Concepts and Techniques. Morgan Kaufmann Publishers, San Francisco, 2001. [22] D. J. Hand. Data Mining: Statistics and More? The American Statistician, 52(2): 112–118, 1998. [23] D. J. Hand, H. Mannila, and P. Smyth. Principles of Data Mining. MIT Press, 2001. [24] T. Hastie, R. Tibshirani, and J. H. Friedman. The Elements of Statistical Learning: Data Mining, Inference, Prediction. Springer, New York, 2001. [25] M. Kantardzic. Data Mining: Concepts, Models, Methods, and Algorithms. WileyIEEE Press, Piscataway, NJ, 2003.
15
Chapter 1
Introduction
[26] H. Kargupta and P. K. Chan, editors. Advances in Distributed and Parallel Knowledge Discovery. AAAI Press, September 2002. [27] H. Kargupta, S. Datta, Q. Wang, and K. Sivakumar. On the Privacy Preserving Properties of Random Data Perturbation Techniques. In Proc. of the 2003 IEEE Intl. Conf. on Data Mining, pages 99–106, Melbourne, Florida, December 2003. IEEE Computer Society. [28] D. Kifer, S. BenDavid, and J. Gehrke. Detecting Change in Data Streams. In Proc. of the 30th VLDB Conf., pages 180–191, Toronto, Canada, 2004. Morgan Kaufmann. [29] D. Lambert. What Use is Statistics for Massive Data? In ACM SIGMOD Workshop on Research Issues in Data Mining and Knowledge Discovery, pages 54–62, 2000. [30] M. H. C. Law, N. Zhang, and A. K. Jain. Nonlinear Manifold Learning for Data Streams. In Proc. of the SIAM Intl. Conf. on Data Mining, Lake Buena Vista, Florida, April 2004. SIAM. [31] T. Mitchell. Machine Learning. McGrawHill, Boston, MA, 1997. [32] S. Papadimitriou, A. Brockwell, and C. Faloutsos. Adaptive, unsupervised stream mining. VLDB Journal, 13(3):222–239, 2004. [33] O. Parr Rud. Data Mining Cookbook: Modeling Data for Marketing, Risk and Customer Relationship Management. John Wiley & Sons, New York, NY, 2001. [34] D. Pyle. Business Modeling and Data Mining. Morgan Kaufmann, San Francisco, CA, 2003. [35] N. Ramakrishnan and A. Grama. Data Mining: From Serendipity to Science—Guest Editors’ Introduction. IEEE Computer, 32(8):34–37, 1999. [36] R. Roiger and M. Geatz. Data Mining: A Tutorial Based Primer. AddisonWesley, 2002. [37] P. Smyth. Breaking out of the BlackBox: Research Challenges in Data Mining. In Proc. of the 2001 ACM SIGMOD Workshop on Research Issues in Data Mining and Knowledge Discovery, 2001. [38] P. Smyth, D. Pregibon, and C. Faloutsos. Datadriven evolution of data mining algorithms. Communications of the ACM, 45(8):33–37, 2002. [39] V. S. Verykios, E. Bertino, I. N. Fovino, L. P. Provenza, Y. Saygin, and Y. Theodoridis. Stateoftheart in privacy preserving data mining. SIGMOD Record, 33(1):50–57, 2004. [40] J. T. L. Wang, M. J. Zaki, H. Toivonen, and D. E. Shasha, editors. Data Mining in Bioinformatics. Springer, September 2004. [41] A. R. Webb. Statistical Pattern Recognition. John Wiley & Sons, 2nd edition, 2002. [42] I. H. Witten and E. Frank. Data Mining: Practical Machine Learning Tools and Techniques with Java Implementations. Morgan Kaufmann, 1999. [43] X. Wu, P. S. Yu, and G. PiatetskyShapiro. Data Mining: How Research Meets Practical Development? Knowledge and Information Systems, 5(2):248–261, 2003. [44] M. J. Zaki and C.T. Ho, editors. LargeScale Parallel Data Mining. Springer, September 2002.
1.7
Exercises
1. Discuss whether or not each of the following activities is a data mining task.
16
1.7
Exercises
(a) Dividing the customers of a company according to their gender. (b) Dividing the customers of a company according to their proﬁtability. (c) Computing the total sales of a company. (d) Sorting a student database based on student identiﬁcation numbers. (e) Predicting the outcomes of tossing a (fair) pair of dice. (f) Predicting the future stock price of a company using historical records. (g) Monitoring the heart rate of a patient for abnormalities. (h) Monitoring seismic waves for earthquake activities. (i) Extracting the frequencies of a sound wave. 2. Suppose that you are employed as a data mining consultant for an Internet search engine company. Describe how data mining can help the company by giving speciﬁc examples of how techniques, such as clustering, classiﬁcation, association rule mining, and anomaly detection can be applied. 3. For each of the following data sets, explain whether or not data privacy is an important issue. (a) Census data collected from 1900–1950. (b) IP addresses and visit times of Web users who visit your Website. (c) Images from Earthorbiting satellites. (d) Names and addresses of people from the telephone book. (e) Names and email addresses collected from the Web.
17
18
2 Data This chapter discusses several datarelated issues that are important for successful data mining: The Type of Data Data sets diﬀer in a number of ways. For example, the attributes used to describe data objects can be of diﬀerent types—quantitative or qualitative—and data sets may have special characteristics; e.g., some data sets contain time series or objects with explicit relationships to one another. Not surprisingly, the type of data determines which tools and techniques can be used to analyze the data. Furthermore, new research in data mining is often driven by the need to accommodate new application areas and their new types of data. The Quality of the Data Data is often far from perfect. While most data mining techniques can tolerate some level of imperfection in the data, a focus on understanding and improving data quality typically improves the quality of the resulting analysis. Data quality issues that often need to be addressed include the presence of noise and outliers; missing, inconsistent, or duplicate data; and data that is biased or, in some other way, unrepresentative of the phenomenon or population that the data is supposed to describe. Preprocessing Steps to Make the Data More Suitable for Data Mining Often, the raw data must be processed in order to make it suitable for analysis. While one objective may be to improve data quality, other goals focus on modifying the data so that it better ﬁts a speciﬁed data mining technique or tool. For example, a continuous attribute, e.g., length, may need to be transformed into an attribute with discrete categories, e.g., short, medium, or long, in order to apply a particular technique. As another example, the
From Chapter 2 of Introduction to Data Mining, First Edition. PangNing Tan, Michael Steinbach, Vipin Kumar. Copyright © 2006 by Pearson Education, Inc. All rights reserved.
19
Chapter 2
Data
number of attributes in a data set is often reduced because many techniques are more eﬀective when the data has a relatively small number of attributes. Analyzing Data in Terms of Its Relationships One approach to data analysis is to ﬁnd relationships among the data objects and then perform the remaining analysis using these relationships rather than the data objects themselves. For instance, we can compute the similarity or distance between pairs of objects and then perform the analysis—clustering, classiﬁcation, or anomaly detection—based on these similarities or distances. There are many such similarity or distance measures, and the proper choice depends on the type of data and the particular application. Example 2.1 (An Illustration of DataRelated Issues). To further illustrate the importance of these issues, consider the following hypothetical situation. You receive an email from a medical researcher concerning a project that you are eager to work on. Hi, I’ve attached the data ﬁle that I mentioned in my previous email. Each line contains the information for a single patient and consists of ﬁve ﬁelds. We want to predict the last ﬁeld using the other ﬁelds. I don’t have time to provide any more information about the data since I’m going out of town for a couple of days, but hopefully that won’t slow you down too much. And if you don’t mind, could we meet when I get back to discuss your preliminary results? I might invite a few other members of my team. Thanks and see you in a couple of days. Despite some misgivings, you proceed to analyze the data. The ﬁrst few rows of the ﬁle are as follows: 012 020 027 .. .
232 121 165
33.5 16.9 24.0
0 2 0
10.7 210.1 427.6
A brief look at the data reveals nothing strange. You put your doubts aside and start the analysis. There are only 1000 lines, a smaller data ﬁle than you had hoped for, but two days later, you feel that you have made some progress. You arrive for the meeting, and while waiting for others to arrive, you strike
20
up a conversation with a statistician who is working on the project. When she learns that you have also been analyzing the data from the project, she asks if you would mind giving her a brief overview of your results. Statistician: So, you got the data for all the patients? Data Miner: Yes. I haven’t had much time for analysis, but I do have a few interesting results. Statistician: Amazing. There were so many data issues with this set of patients that I couldn’t do much. Data Miner: Oh? I didn’t hear about any possible problems. Statistician: Well, ﬁrst there is ﬁeld 5, the variable we want to predict. It’s common knowledge among people who analyze this type of data that results are better if you work with the log of the values, but I didn’t discover this until later. Was it mentioned to you? Data Miner: No. Statistician: But surely you heard about what happened to ﬁeld 4? It’s supposed to be measured on a scale from 1 to 10, with 0 indicating a missing value, but because of a data entry error, all 10’s were changed into 0’s. Unfortunately, since some of the patients have missing values for this ﬁeld, it’s impossible to say whether a 0 in this ﬁeld is a real 0 or a 10. Quite a few of the records have that problem. Data Miner: Interesting. Were there any other problems? Statistician: Yes, ﬁelds 2 and 3 are basically the same, but I assume that you probably noticed that. Data Miner: Yes, but these ﬁelds were only weak predictors of ﬁeld 5. Statistician: Anyway, given all those problems, I’m surprised you were able to accomplish anything. Data Miner: True, but my results are really quite good. Field 1 is a very strong predictor of ﬁeld 5. I’m surprised that this wasn’t noticed before. Statistician: What? Field 1 is just an identiﬁcation number. Data Miner: Nonetheless, my results speak for themselves. Statistician: Oh, no! I just remembered. We assigned ID numbers after we sorted the records based on ﬁeld 5. There is a strong connection, but it’s meaningless. Sorry.
21
Chapter 2
Data
Although this scenario represents an extreme situation, it emphasizes the importance of “knowing your data.” To that end, this chapter will address each of the four issues mentioned above, outlining some of the basic challenges and standard approaches.
2.1
Types of Data
A data set can often be viewed as a collection of data objects. Other names for a data object are record, point, vector, pattern, event, case, sample, observation, or entity. In turn, data objects are described by a number of attributes that capture the basic characteristics of an object, such as the mass of a physical object or the time at which an event occurred. Other names for an attribute are variable, characteristic, ﬁeld, feature, or dimension. Example 2.2 (Student Information). Often, a data set is a ﬁle, in which the objects are records (or rows) in the ﬁle and each ﬁeld (or column) corresponds to an attribute. For example, Table 2.1 shows a data set that consists of student information. Each row corresponds to a student and each column is an attribute that describes some aspect of a student, such as grade point average (GPA) or identiﬁcation number (ID). Table 2.1. A sample data set containing student information. Student ID 1034262 1052663 1082246
Year .. . Senior Sophomore Freshman .. .
Grade Point Average (GPA)
...
3.24 3.51 3.62
... ... ...
Although recordbased data sets are common, either in ﬂat ﬁles or relational database systems, there are other important types of data sets and systems for storing data. In Section 2.1.2, we will discuss some of the types of data sets that are commonly encountered in data mining. However, we ﬁrst consider attributes.
22
2.1
2.1.1
Types of Data
Attributes and Measurement
In this section we address the issue of describing data by considering what types of attributes are used to describe data objects. We ﬁrst deﬁne an attribute, then consider what we mean by the type of an attribute, and ﬁnally describe the types of attributes that are commonly encountered. What Is an attribute? We start with a more detailed deﬁnition of an attribute. Deﬁnition 2.1. An attribute is a property or characteristic of an object that may vary, either from one object to another or from one time to another. For example, eye color varies from person to person, while the temperature of an object varies over time. Note that eye color is a symbolic attribute with a small number of possible values {brown, black, blue, green, hazel, etc.}, while temperature is a numerical attribute with a potentially unlimited number of values. At the most basic level, attributes are not about numbers or symbols. However, to discuss and more precisely analyze the characteristics of objects, we assign numbers or symbols to them. To do this in a welldeﬁned way, we need a measurement scale. Deﬁnition 2.2. A measurement scale is a rule (function) that associates a numerical or symbolic value with an attribute of an object. Formally, the process of measurement is the application of a measurement scale to associate a value with a particular attribute of a speciﬁc object. While this may seem a bit abstract, we engage in the process of measurement all the time. For instance, we step on a bathroom scale to determine our weight, we classify someone as male or female, or we count the number of chairs in a room to see if there will be enough to seat all the people coming to a meeting. In all these cases, the “physical value” of an attribute of an object is mapped to a numerical or symbolic value. With this background, we can now discuss the type of an attribute, a concept that is important in determining if a particular data analysis technique is consistent with a speciﬁc type of attribute. The Type of an Attribute It should be apparent from the previous discussion that the properties of an attribute need not be the same as the properties of the values used to mea
23
Chapter 2
Data
sure it. In other words, the values used to represent an attribute may have properties that are not properties of the attribute itself, and vice versa. This is illustrated with two examples. Example 2.3 (Employee Age and ID Number). Two attributes that might be associated with an employee are ID and age (in years). Both of these attributes can be represented as integers. However, while it is reasonable to talk about the average age of an employee, it makes no sense to talk about the average employee ID. Indeed, the only aspect of employees that we want to capture with the ID attribute is that they are distinct. Consequently, the only valid operation for employee IDs is to test whether they are equal. There is no hint of this limitation, however, when integers are used to represent the employee ID attribute. For the age attribute, the properties of the integers used to represent age are very much the properties of the attribute. Even so, the correspondence is not complete since, for example, ages have a maximum, while integers do not. Example 2.4 (Length of Line Segments). Consider Figure 2.1, which shows some objects—line segments—and how the length attribute of these objects can be mapped to numbers in two diﬀerent ways. Each successive line segment, going from the top to the bottom, is formed by appending the topmost line segment to itself. Thus, the second line segment from the top is formed by appending the topmost line segment to itself twice, the third line segment from the top is formed by appending the topmost line segment to itself three times, and so forth. In a very real (physical) sense, all the line segments are multiples of the ﬁrst. This fact is captured by the measurements on the righthand side of the ﬁgure, but not by those on the left handside. More speciﬁcally, the measurement scale on the lefthand side captures only the ordering of the length attribute, while the scale on the righthand side captures both the ordering and additivity properties. Thus, an attribute can be measured in a way that does not capture all the properties of the attribute. The type of an attribute should tell us what properties of the attribute are reﬂected in the values used to measure it. Knowing the type of an attribute is important because it tells us which properties of the measured values are consistent with the underlying properties of the attribute, and therefore, it allows us to avoid foolish actions, such as computing the average employee ID. Note that it is common to refer to the type of an attribute as the type of a measurement scale.
24
2.1
Types of Data
1
1
3
2
7
3
8
4
10
5
A mapping of lengths to numbers that captures only the order properties of length.
A mapping of lengths to numbers that captures both the order and additivity properties of length.
Figure 2.1. The measurement of the length of line segments on two different scales of measurement.
The Diﬀerent Types of Attributes A useful (and simple) way to specify the type of an attribute is to identify the properties of numbers that correspond to underlying properties of the attribute. For example, an attribute such as length has many of the properties of numbers. It makes sense to compare and order objects by length, as well as to talk about the diﬀerences and ratios of length. The following properties (operations) of numbers are typically used to describe attributes. 1. Distinctness = and = 2. Order , and ≥ 3. Addition + and − 4. Multiplication ∗ and / Given these properties, we can deﬁne four types of attributes: nominal, ordinal, interval, and ratio. Table 2.2 gives the deﬁnitions of these types, along with information about the statistical operations that are valid for each type. Each attribute type possesses all of the properties and operations of the attribute types above it. Consequently, any property or operation that is valid for nominal, ordinal, and interval attributes is also valid for ratio attributes. In other words, the deﬁnition of the attribute types is cumulative. However,
25
Chapter 2
Data Table 2.2. Different attribute types.
Numeric (Quantitative)
Categorical (Qualitative)
Attribute Description Type Nominal The values of a nominal attribute are just diﬀerent names; i.e., nominal values provide only enough information to distinguish one object from another. (=, =) Ordinal The values of an ordinal attribute provide enough information to order objects. () Interval For interval attributes, the diﬀerences between values are meaningful, i.e., a unit of measurement exists. (+, − ) Ratio For ratio variables, both diﬀerences and ratios are meaningful. (*, /)
Examples zip codes, employee ID numbers, eye color, gender
Operations mode, entropy, contingency correlation, χ2 test
hardness of minerals, {good, better, best}, grades, street numbers
median, percentiles, rank correlation, run tests, sign tests mean, standard deviation, Pearson’s correlation, t and F tests geometric mean, harmonic mean, percent variation
calendar dates, temperature in Celsius or Fahrenheit
temperature in Kelvin, monetary quantities, counts, age, mass, length, electrical current
this does not mean that the operations appropriate for one attribute type are appropriate for the attribute types above it. Nominal and ordinal attributes are collectively referred to as categorical or qualitative attributes. As the name suggests, qualitative attributes, such as employee ID, lack most of the properties of numbers. Even if they are represented by numbers, i.e., integers, they should be treated more like symbols. The remaining two types of attributes, interval and ratio, are collectively referred to as quantitative or numeric attributes. Quantitative attributes are represented by numbers and have most of the properties of numbers. Note that quantitative attributes can be integervalued or continuous. The types of attributes can also be described in terms of transformations that do not change the meaning of an attribute. Indeed, S. Smith Stevens, the psychologist who originally deﬁned the types of attributes shown in Table 2.2, deﬁned them in terms of these permissible transformations. For example,
26
2.1
Types of Data
Table 2.3. Transformations that define attribute levels.
Numeric (Quantitative)
Categorical (Qualitative)
Attribute Type Nominal
Transformation Any onetoone mapping, e.g., a permutation of values
Ordinal
An orderpreserving change of values, i.e., new value = f (old value), where f is a monotonic function.
Interval
new value = a ∗ old value + b, a and b constants.
Ratio
new value = a ∗ old value
Comment If all employee ID numbers are reassigned, it will not make any diﬀerence. An attribute encompassing the notion of good, better, best can be represented equally well by the values {1, 2, 3} or by {0.5, 1, 10}. The Fahrenheit and Celsius temperature scales diﬀer in the location of their zero value and the size of a degree (unit). Length can be measured in meters or feet.
the meaning of a length attribute is unchanged if it is measured in meters instead of feet. The statistical operations that make sense for a particular type of attribute are those that will yield the same results when the attribute is transformed using a transformation that preserves the attribute’s meaning. To illustrate, the average length of a set of objects is diﬀerent when measured in meters rather than in feet, but both averages represent the same length. Table 2.3 shows the permissible (meaningpreserving) transformations for the four attribute types of Table 2.2. Example 2.5 (Temperature Scales). Temperature provides a good illustration of some of the concepts that have been described. First, temperature can be either an interval or a ratio attribute, depending on its measurement scale. When measured on the Kelvin scale, a temperature of 2◦ is, in a physically meaningful way, twice that of a temperature of 1◦ . This is not true when temperature is measured on either the Celsius or Fahrenheit scales, because, physically, a temperature of 1◦ Fahrenheit (Celsius) is not much diﬀerent than a temperature of 2◦ Fahrenheit (Celsius). The problem is that the zero points of the Fahrenheit and Celsius scales are, in a physical sense, arbitrary, and therefore, the ratio of two Celsius or Fahrenheit temperatures is not physically meaningful.
27
Chapter 2
Data
Describing Attributes by the Number of Values An independent way of distinguishing between attributes is by the number of values they can take. Discrete A discrete attribute has a ﬁnite or countably inﬁnite set of values. Such attributes can be categorical, such as zip codes or ID numbers, or numeric, such as counts. Discrete attributes are often represented using integer variables. Binary attributes are a special case of discrete attributes and assume only two values, e.g., true/false, yes/no, male/female, or 0/1. Binary attributes are often represented as Boolean variables, or as integer variables that only take the values 0 or 1. Continuous A continuous attribute is one whose values are real numbers. Examples include attributes such as temperature, height, or weight. Continuous attributes are typically represented as ﬂoatingpoint variables. Practically, real values can only be measured and represented with limited precision. In theory, any of the measurement scale types—nominal, ordinal, interval, and ratio—could be combined with any of the types based on the number of attribute values—binary, discrete, and continuous. However, some combinations occur only infrequently or do not make much sense. For instance, it is diﬃcult to think of a realistic data set that contains a continuous binary attribute. Typically, nominal and ordinal attributes are binary or discrete, while interval and ratio attributes are continuous. However, count attributes, which are discrete, are also ratio attributes. Asymmetric Attributes For asymmetric attributes, only presence—a nonzero attribute value—is regarded as important. Consider a data set where each object is a student and each attribute records whether or not a student took a particular course at a university. For a speciﬁc student, an attribute has a value of 1 if the student took the course associated with that attribute and a value of 0 otherwise. Because students take only a small fraction of all available courses, most of the values in such a data set would be 0. Therefore, it is more meaningful and more eﬃcient to focus on the nonzero values. To illustrate, if students are compared on the basis of the courses they don’t take, then most students would seem very similar, at least if the number of courses is large. Binary attributes where only nonzero values are important are called asymmetric
28
2.1
Types of Data
binary attributes. This type of attribute is particularly important for association analysis, which is discussed in Chapter 6. It is also possible to have discrete or continuous asymmetric features. For instance, if the number of credits associated with each course is recorded, then the resulting data set will consist of asymmetric discrete or continuous attributes.
2.1.2
Types of Data Sets
There are many types of data sets, and as the ﬁeld of data mining develops and matures, a greater variety of data sets become available for analysis. In this section, we describe some of the most common types. For convenience, we have grouped the types of data sets into three groups: record data, graphbased data, and ordered data. These categories do not cover all possibilities and other groupings are certainly possible. General Characteristics of Data Sets Before providing details of speciﬁc kinds of data sets, we discuss three characteristics that apply to many data sets and have a signiﬁcant impact on the data mining techniques that are used: dimensionality, sparsity, and resolution. Dimensionality The dimensionality of a data set is the number of attributes that the objects in the data set possess. Data with a small number of dimensions tends to be qualitatively diﬀerent than moderate or highdimensional data. Indeed, the diﬃculties associated with analyzing highdimensional data are sometimes referred to as the curse of dimensionality. Because of this, an important motivation in preprocessing the data is dimensionality reduction. These issues are discussed in more depth later in this chapter and in Appendix B. Sparsity For some data sets, such as those with asymmetric features, most attributes of an object have values of 0; in many cases, fewer than 1% of the entries are nonzero. In practical terms, sparsity is an advantage because usually only the nonzero values need to be stored and manipulated. This results in signiﬁcant savings with respect to computation time and storage. Furthermore, some data mining algorithms work well only for sparse data. Resolution It is frequently possible to obtain data at diﬀerent levels of resolution, and often the properties of the data are diﬀerent at diﬀerent resolutions. For instance, the surface of the Earth seems very uneven at a resolution of a
29
Chapter 2
Data
few meters, but is relatively smooth at a resolution of tens of kilometers. The patterns in the data also depend on the level of resolution. If the resolution is too ﬁne, a pattern may not be visible or may be buried in noise; if the resolution is too coarse, the pattern may disappear. For example, variations in atmospheric pressure on a scale of hours reﬂect the movement of storms and other weather systems. On a scale of months, such phenomena are not detectable. Record Data Much data mining work assumes that the data set is a collection of records (data objects), each of which consists of a ﬁxed set of data ﬁelds (attributes). See Figure 2.2(a). For the most basic form of record data, there is no explicit relationship among records or data ﬁelds, and every record (object) has the same set of attributes. Record data is usually stored either in ﬂat ﬁles or in relational databases. Relational databases are certainly more than a collection of records, but data mining often does not use any of the additional information available in a relational database. Rather, the database serves as a convenient place to ﬁnd records. Diﬀerent types of record data are described below and are illustrated in Figure 2.2. Transaction or Market Basket Data Transaction data is a special type of record data, where each record (transaction) involves a set of items. Consider a grocery store. The set of products purchased by a customer during one shopping trip constitutes a transaction, while the individual products that were purchased are the items. This type of data is called market basket data because the items in each record are the products in a person’s “market basket.” Transaction data is a collection of sets of items, but it can be viewed as a set of records whose ﬁelds are asymmetric attributes. Most often, the attributes are binary, indicating whether or not an item was purchased, but more generally, the attributes can be discrete or continuous, such as the number of items purchased or the amount spent on those items. Figure 2.2(b) shows a sample transaction data set. Each row represents the purchases of a particular customer at a particular time. The Data Matrix If the data objects in a collection of data all have the same ﬁxed set of numeric attributes, then the data objects can be thought of as points (vectors) in a multidimensional space, where each dimension represents a distinct attribute describing the object. A set of such data objects can be interpreted as an m by n matrix, where there are m rows, one for each object,
30
2.1 Tid Refund Marital Status
Taxable Defaulted Income Borrower
1
Yes
Single
125K
No
2
No
Married
100K
No
3
No
Single
70K
No
4
Yes
Married
120K
No
5
No
Divorced 95K
Yes
6
No
Married
No
7
Yes
Divorced 220K
No
8
No
Single
85K
Yes
9
No
Married
75K
No
10
No
Single
90K
Yes
60K
TID
ITEMS
1
Bread, Soda, Milk
2
Beer, Bread
3
Beer, Soda, Diaper, Milk
4
Beer, Bread, Diaper, Milk
5
Soda, Diaper, Milk
(a) Record data.
(b) Transaction data.
team
play
ball
score
game
win
lost
timeout
season
Document 1
3
0
5
0
2
6
0
2
0
2
Document 2
0
7
0
2
1
0
0
3
0
0
Document 3
0
1
0
0
1
2
2
0
3
0
Thickness
10.23
5.27
15.22
27
1.2
12.65
6.25
16.22
22
1.1
13.54
7.23
17.34
23
1.2
14.27
8.43
18.45
25
0.9
(c) Data matrix.
coach
Projection of Projection of Distance Load x Load y Load
Types of Data
(d) Documentterm matrix.
Figure 2.2. Different variations of record data.
and n columns, one for each attribute. (A representation that has data objects as columns and attributes as rows is also ﬁne.) This matrix is called a data matrix or a pattern matrix. A data matrix is a variation of record data, but because it consists of numeric attributes, standard matrix operation can be applied to transform and manipulate the data. Therefore, the data matrix is the standard data format for most statistical data. Figure 2.2(c) shows a sample data matrix. The Sparse Data Matrix A sparse data matrix is a special case of a data matrix in which the attributes are of the same type and are asymmetric; i.e., only nonzero values are important. Transaction data is an example of a sparse data matrix that has only 0–1 entries. Another common example is document data. In particular, if the order of the terms (words) in a document is ignored,
31
Chapter 2
Data
then a document can be represented as a term vector, where each term is a component (attribute) of the vector and the value of each component is the number of times the corresponding term occurs in the document. This representation of a collection of documents is often called a documentterm matrix. Figure 2.2(d) shows a sample documentterm matrix. The documents are the rows of this matrix, while the terms are the columns. In practice, only the nonzero entries of sparse data matrices are stored. GraphBased Data A graph can sometimes be a convenient and powerful representation for data. We consider two speciﬁc cases: (1) the graph captures relationships among data objects and (2) the data objects themselves are represented as graphs. Data with Relationships among Objects The relationships among objects frequently convey important information. In such cases, the data is often represented as a graph. In particular, the data objects are mapped to nodes of the graph, while the relationships among objects are captured by the links between objects and link properties, such as direction and weight. Consider Web pages on the World Wide Web, which contain both text and links to other pages. In order to process search queries, Web search engines collect and process Web pages to extract their contents. It is well known, however, that the links to and from each page provide a great deal of information about the relevance of a Web page to a query, and thus, must also be taken into consideration. Figure 2.3(a) shows a set of linked Web pages. Data with Objects That Are Graphs If objects have structure, that is, the objects contain subobjects that have relationships, then such objects are frequently represented as graphs. For example, the structure of chemical compounds can be represented by a graph, where the nodes are atoms and the links between nodes are chemical bonds. Figure 2.3(b) shows a ballandstick diagram of the chemical compound benzene, which contains atoms of carbon (black) and hydrogen (gray). A graph representation makes it possible to determine which substructures occur frequently in a set of compounds and to ascertain whether the presence of any of these substructures is associated with the presence or absence of certain chemical properties, such as melting point or heat of formation. Substructure mining, which is a branch of data mining that analyzes such data, is considered in Section 7.5.
32
2.1
Types of Data
Useful Links: •
Bibliography
•
Other Useful Web sites
Knowledge Discovery and Data Mining Bibliography (Gets updated frequently, so visit often!)
o
ACM SIGKDD
o
KDnuggets
•
Books
o
The Data Mine
•
General Data Mining
Book References in Data Mining and Knowledge Discovery Usama Fayyad, Gregory PiatetskyShapiro, Padhraic Smyth, and Ramasamy uthurasamy, "Advances in Knowledge Discovery and Data Mining", AAAI Press/the MIT Press, 1996. J. Ross Quinlan, "C4.5: Programs for Machine Learning", Morgan Kaufmann Publishers, 1993. Michael Berry and Gordon Linoff, "Data Mining Techniques (For Marketing, Sales, and Customer Support), John Wiley & Sons, 1997.
General Data Mining Usama Fayyad, "Mining Databases: Towards Algorithms for Knowledge Discovery", Bulletin of the IEEE Computer Society Technical Committee on data Engineering, vol. 21, no. 1, March 1998. Christopher Matheus, Philip Chan, and Gregory PiatetskyShapiro, "Systems for knowledge Discovery in databases", IEEE Transactions on Knowledge and Data Engineering, 5(6):903913, December 1993.
(a) Linked Web pages.
(b) Benzene molecule.
Figure 2.3. Different variations of graph data.
Ordered Data For some types of data, the attributes have relationships that involve order in time or space. Diﬀerent types of ordered data are described next and are shown in Figure 2.4. Sequential Data Sequential data, also referred to as temporal data, can be thought of as an extension of record data, where each record has a time associated with it. Consider a retail transaction data set that also stores the time at which the transaction took place. This time information makes it possible to ﬁnd patterns such as “candy sales peak before Halloween.” A time can also be associated with each attribute. For example, each record could be the purchase history of a customer, with a listing of items purchased at diﬀerent times. Using this information, it is possible to ﬁnd patterns such as “people who buy DVD players tend to buy DVDs in the period immediately following the purchase.” Figure 2.4(a) shows an example of sequential transaction data. There are ﬁve diﬀerent times—t1, t2, t3, t4, and t5 ; three diﬀerent customers—C1,
33
Chapter 2
Time t1 t2 t2 t3 t4 t5
Customer C1 C3 C1 C2 C2 C1
Customer C1 C2 C3
Data
GGTTCCGCCTTCAGCCCCGCGCC CGCAGGGCCCGCCCCGCGCCGTC GAGAAGGGCCCGCCTGGCGGGCG GGGGGAGGCGGGGCCGCCCGAGC CCAACCGAGTCCGACCAGGTGCC CCCTCTGCTCGGCCTAGACCTGA GCTCATTAGGCGGCAGCGGACAG GCCAAGTAGAACACGCGAAGCGC TGGGCTGCCTGCTGCGACCAGGG
Items Purchased A, B A, C C, D A, D E A, E
Time and Items Purchased (t1: A,B) (t2:C,D) (t5:A,E) (t3: A, D) (t4: E) (t2: A, C)
(a) Sequential transaction data.
(b) Genomic sequence data.
Minneapolis Average Monthly Temperature (1982–1993) 30
90 30
25
60 20
30 20
10
Latitude
Temperature (celcius)
25 15
5
0
15
–30
10
0 –5
5
–10
–60
–15
0
–20 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 Year
(c) Temperature time series.
–90 –180 –150 –120 –90
–60
–30
0 30 Longitude
60
90
120
150 180
Temp
(d) Spatial temperature data.
Figure 2.4. Different variations of ordered data.
C2, and C3; and ﬁve diﬀerent items—A, B, C, D, and E. In the top table, each row corresponds to the items purchased at a particular time by each customer. For instance, at time t3, customer C2 purchased items A and D. In the bottom table, the same information is displayed, but each row corresponds to a particular customer. Each row contains information on each transaction involving the customer, where a transaction is considered to be a set of items and the time at which those items were purchased. For example, customer C3 bought items A and C at time t2.
34
2.1
Types of Data
Sequence Data Sequence data consists of a data set that is a sequence of individual entities, such as a sequence of words or letters. It is quite similar to sequential data, except that there are no time stamps; instead, there are positions in an ordered sequence. For example, the genetic information of plants and animals can be represented in the form of sequences of nucleotides that are known as genes. Many of the problems associated with genetic sequence data involve predicting similarities in the structure and function of genes from similarities in nucleotide sequences. Figure 2.4(b) shows a section of the human genetic code expressed using the four nucleotides from which all DNA is constructed: A, T, G, and C. Time Series Data Time series data is a special type of sequential data in which each record is a time series, i.e., a series of measurements taken over time. For example, a ﬁnancial data set might contain objects that are time series of the daily prices of various stocks. As another example, consider Figure 2.4(c), which shows a time series of the average monthly temperature for Minneapolis during the years 1982 to 1994. When working with temporal data, it is important to consider temporal autocorrelation; i.e., if two measurements are close in time, then the values of those measurements are often very similar. Spatial Data Some objects have spatial attributes, such as positions or areas, as well as other types of attributes. An example of spatial data is weather data (precipitation, temperature, pressure) that is collected for a variety of geographical locations. An important aspect of spatial data is spatial autocorrelation; i.e., objects that are physically close tend to be similar in other ways as well. Thus, two points on the Earth that are close to each other usually have similar values for temperature and rainfall. Important examples of spatial data are the science and engineering data sets that are the result of measurements or model output taken at regularly or irregularly distributed points on a two or threedimensional grid or mesh. For instance, Earth science data sets record the temperature or pressure measured at points (grid cells) on latitude–longitude spherical grids of various resolutions, e.g., 1◦ by 1◦ . (See Figure 2.4(d).) As another example, in the simulation of the ﬂow of a gas, the speed and direction of ﬂow can be recorded for each grid point in the simulation.
35
Chapter 2
Data
Handling NonRecord Data Most data mining algorithms are designed for record data or its variations, such as transaction data and data matrices. Recordoriented techniques can be applied to nonrecord data by extracting features from data objects and using these features to create a record corresponding to each object. Consider the chemical structure data that was described earlier. Given a set of common substructures, each compound can be represented as a record with binary attributes that indicate whether a compound contains a speciﬁc substructure. Such a representation is actually a transaction data set, where the transactions are the compounds and the items are the substructures. In some cases, it is easy to represent the data in a record format, but this type of representation does not capture all the information in the data. Consider spatiotemporal data consisting of a time series from each point on a spatial grid. This data is often stored in a data matrix, where each row represents a location and each column represents a particular point in time. However, such a representation does not explicitly capture the time relationships that are present among attributes and the spatial relationships that exist among objects. This does not mean that such a representation is inappropriate, but rather that these relationships must be taken into consideration during the analysis. For example, it would not be a good idea to use a data mining technique that assumes the attributes are statistically independent of one another.
2.2
Data Quality
Data mining applications are often applied to data that was collected for another purpose, or for future, but unspeciﬁed applications. For that reason, data mining cannot usually take advantage of the signiﬁcant beneﬁts of “addressing quality issues at the source.” In contrast, much of statistics deals with the design of experiments or surveys that achieve a prespeciﬁed level of data quality. Because preventing data quality problems is typically not an option, data mining focuses on (1) the detection and correction of data quality problems and (2) the use of algorithms that can tolerate poor data quality. The ﬁrst step, detection and correction, is often called data cleaning. The following sections discuss speciﬁc aspects of data quality. The focus is on measurement and data collection issues, although some applicationrelated issues are also discussed.
36
2.2
2.2.1
Data Quality
Measurement and Data Collection Issues
It is unrealistic to expect that data will be perfect. There may be problems due to human error, limitations of measuring devices, or ﬂaws in the data collection process. Values or even entire data objects may be missing. In other cases, there may be spurious or duplicate objects; i.e., multiple data objects that all correspond to a single “real” object. For example, there might be two diﬀerent records for a person who has recently lived at two diﬀerent addresses. Even if all the data is present and “looks ﬁne,” there may be inconsistencies—a person has a height of 2 meters, but weighs only 2 kilograms. In the next few sections, we focus on aspects of data quality that are related to data measurement and collection. We begin with a deﬁnition of measurement and data collection errors and then consider a variety of problems that involve measurement error: noise, artifacts, bias, precision, and accuracy. We conclude by discussing data quality issues that may involve both measurement and data collection problems: outliers, missing and inconsistent values, and duplicate data. Measurement and Data Collection Errors The term measurement error refers to any problem resulting from the measurement process. A common problem is that the value recorded diﬀers from the true value to some extent. For continuous attributes, the numerical difference of the measured and true value is called the error. The term data collection error refers to errors such as omitting data objects or attribute values, or inappropriately including a data object. For example, a study of animals of a certain species might include animals of a related species that are similar in appearance to the species of interest. Both measurement errors and data collection errors can be either systematic or random. We will only consider general types of errors. Within particular domains, there are certain types of data errors that are commonplace, and there often exist welldeveloped techniques for detecting and/or correcting these errors. For example, keyboard errors are common when data is entered manually, and as a result, many data entry programs have techniques for detecting and, with human intervention, correcting such errors. Noise and Artifacts Noise is the random component of a measurement error. It may involve the distortion of a value or the addition of spurious objects. Figure 2.5 shows a time series before and after it has been disrupted by random noise. If a bit
37
Chapter 2
Data
(a) Time series.
(b) Time series with noise.
Figure 2.5. Noise in a time series context.
(a) Three groups of points.
(b) With noise points (+) added.
Figure 2.6. Noise in a spatial context.
more noise were added to the time series, its shape would be lost. Figure 2.6 shows a set of data points before and after some noise points (indicated by ‘+’s) have been added. Notice that some of the noise points are intermixed with the nonnoise points. The term noise is often used in connection with data that has a spatial or temporal component. In such cases, techniques from signal or image processing can frequently be used to reduce noise and thus, help to discover patterns (signals) that might be “lost in the noise.” Nonetheless, the elimination of noise is frequently diﬃcult, and much work in data mining focuses on devising robust algorithms that produce acceptable results even when noise is present.
38
2.2
Data Quality
Data errors may be the result of a more deterministic phenomenon, such as a streak in the same place on a set of photographs. Such deterministic distortions of the data are often referred to as artifacts. Precision, Bias, and Accuracy In statistics and experimental science, the quality of the measurement process and the resulting data are measured by precision and bias. We provide the standard deﬁnitions, followed by a brief discussion. For the following deﬁnitions, we assume that we make repeated measurements of the same underlying quantity and use this set of values to calculate a mean (average) value that serves as our estimate of the true value. Deﬁnition 2.3 (Precision). The closeness of repeated measurements (of the same quantity) to one another. Deﬁnition 2.4 (Bias). A systematic variation of measurements from the quantity being measured. Precision is often measured by the standard deviation of a set of values, while bias is measured by taking the diﬀerence between the mean of the set of values and the known value of the quantity being measured. Bias can only be determined for objects whose measured quantity is known by means external to the current situation. Suppose that we have a standard laboratory weight with a mass of 1g and want to assess the precision and bias of our new laboratory scale. We weigh the mass ﬁve times, and obtain the following ﬁve values: {1.015, 0.990, 1.013, 1.001, 0.986}. The mean of these values is 1.001, and hence, the bias is 0.001. The precision, as measured by the standard deviation, is 0.013. It is common to use the more general term, accuracy, to refer to the degree of measurement error in data. Deﬁnition 2.5 (Accuracy). The closeness of measurements to the true value of the quantity being measured. Accuracy depends on precision and bias, but since it is a general concept, there is no speciﬁc formula for accuracy in terms of these two quantities. One important aspect of accuracy is the use of signiﬁcant digits. The goal is to use only as many digits to represent the result of a measurement or calculation as are justiﬁed by the precision of the data. For example, if the length of an object is measured with a meter stick whose smallest markings are millimeters, then we should only record the length of data to the nearest millimeter. The precision of such a measurement would be ± 0.5mm. We do not
39
Chapter 2
Data
review the details of working with signiﬁcant digits, as most readers will have encountered them in previous courses, and they are covered in considerable depth in science, engineering, and statistics textbooks. Issues such as signiﬁcant digits, precision, bias, and accuracy are sometimes overlooked, but they are important for data mining as well as statistics and science. Many times, data sets do not come with information on the precision of the data, and furthermore, the programs used for analysis return results without any such information. Nonetheless, without some understanding of the accuracy of the data and the results, an analyst runs the risk of committing serious data analysis blunders. Outliers Outliers are either (1) data objects that, in some sense, have characteristics that are diﬀerent from most of the other data objects in the data set, or (2) values of an attribute that are unusual with respect to the typical values for that attribute. Alternatively, we can speak of anomalous objects or values. There is considerable leeway in the deﬁnition of an outlier, and many diﬀerent deﬁnitions have been proposed by the statistics and data mining communities. Furthermore, it is important to distinguish between the notions of noise and outliers. Outliers can be legitimate data objects or values. Thus, unlike noise, outliers may sometimes be of interest. In fraud and network intrusion detection, for example, the goal is to ﬁnd unusual objects or events from among a large number of normal ones. Chapter 10 discusses anomaly detection in more detail. Missing Values It is not unusual for an object to be missing one or more attribute values. In some cases, the information was not collected; e.g., some people decline to give their age or weight. In other cases, some attributes are not applicable to all objects; e.g., often, forms have conditional parts that are ﬁlled out only when a person answers a previous question in a certain way, but for simplicity, all ﬁelds are stored. Regardless, missing values should be taken into account during the data analysis. There are several strategies (and variations on these strategies) for dealing with missing data, each of which may be appropriate in certain circumstances. These strategies are listed next, along with an indication of their advantages and disadvantages.
40
2.2
Data Quality
Eliminate Data Objects or Attributes A simple and eﬀective strategy is to eliminate objects with missing values. However, even a partially speciﬁed data object contains some information, and if many objects have missing values, then a reliable analysis can be diﬃcult or impossible. Nonetheless, if a data set has only a few objects that have missing values, then it may be expedient to omit them. A related strategy is to eliminate attributes that have missing values. This should be done with caution, however, since the eliminated attributes may be the ones that are critical to the analysis. Estimate Missing Values Sometimes missing data can be reliably estimated. For example, consider a time series that changes in a reasonably smooth fashion, but has a few, widely scattered missing values. In such cases, the missing values can be estimated (interpolated) by using the remaining values. As another example, consider a data set that has many similar data points. In this situation, the attribute values of the points closest to the point with the missing value are often used to estimate the missing value. If the attribute is continuous, then the average attribute value of the nearest neighbors is used; if the attribute is categorical, then the most commonly occurring attribute value can be taken. For a concrete illustration, consider precipitation measurements that are recorded by ground stations. For areas not containing a ground station, the precipitation can be estimated using values observed at nearby ground stations. Ignore the Missing Value during Analysis Many data mining approaches can be modiﬁed to ignore missing values. For example, suppose that objects are being clustered and the similarity between pairs of data objects needs to be calculated. If one or both objects of a pair have missing values for some attributes, then the similarity can be calculated by using only the attributes that do not have missing values. It is true that the similarity will only be approximate, but unless the total number of attributes is small or the number of missing values is high, this degree of inaccuracy may not matter much. Likewise, many classiﬁcation schemes can be modiﬁed to work with missing values. Inconsistent Values Data can contain inconsistent values. Consider an address ﬁeld, where both a zip code and city are listed, but the speciﬁed zip code area is not contained in that city. It may be that the individual entering this information transposed two digits, or perhaps a digit was misread when the information was scanned
41
Chapter 2
Data
from a handwritten form. Regardless of the cause of the inconsistent values, it is important to detect and, if possible, correct such problems. Some types of inconsistences are easy to detect. For instance, a person’s height should not be negative. In other cases, it can be necessary to consult an external source of information. For example, when an insurance company processes claims for reimbursement, it checks the names and addresses on the reimbursement forms against a database of its customers. Once an inconsistency has been detected, it is sometimes possible to correct the data. A product code may have “check” digits, or it may be possible to doublecheck a product code against a list of known product codes, and then correct the code if it is incorrect, but close to a known code. The correction of an inconsistency requires additional or redundant information. Example 2.6 (Inconsistent Sea Surface Temperature). This example illustrates an inconsistency in actual time series data that measures the sea surface temperature (SST) at various points on the ocean. SST data was originally collected using oceanbased measurements from ships or buoys, but more recently, satellites have been used to gather the data. To create a longterm data set, both sources of data must be used. However, because the data comes from diﬀerent sources, the two parts of the data are subtly diﬀerent. This discrepancy is visually displayed in Figure 2.7, which shows the correlation of SST values between pairs of years. If a pair of years has a positive correlation, then the location corresponding to the pair of years is colored white; otherwise it is colored black. (Seasonal variations were removed from the data since, otherwise, all the years would be highly correlated.) There is a distinct change in behavior where the data has been put together in 1983. Years within each of the two groups, 1958–1982 and 1983–1999, tend to have a positive correlation with one another, but a negative correlation with years in the other group. This does not mean that this data should not be used, only that the analyst should consider the potential impact of such discrepancies on the data mining analysis. Duplicate Data A data set may include data objects that are duplicates, or almost duplicates, of one another. Many people receive duplicate mailings because they appear in a database multiple times under slightly diﬀerent names. To detect and eliminate such duplicates, two main issues must be addressed. First, if there are two objects that actually represent a single object, then the values of corresponding attributes may diﬀer, and these inconsistent values must be
42
2.2
Data Quality
60 65
Year
70 75 80 85 90 95 60
65
70
75
80
85
90
95
Year
Figure 2.7. Correlation of SST data between pairs of years. White areas indicate positive correlation. Black areas indicate negative correlation.
resolved. Second, care needs to be taken to avoid accidentally combining data objects that are similar, but not duplicates, such as two distinct people with identical names. The term deduplication is often used to refer to the process of dealing with these issues. In some cases, two or more objects are identical with respect to the attributes measured by the database, but they still represent diﬀerent objects. Here, the duplicates are legitimate, but may still cause problems for some algorithms if the possibility of identical objects is not speciﬁcally accounted for in their design. An example of this is given in Exercise 13 on page 91.
2.2.2
Issues Related to Applications
Data quality issues can also be considered from an application viewpoint as expressed by the statement “data is of high quality if it is suitable for its intended use.” This approach to data quality has proven quite useful, particularly in business and industry. A similar viewpoint is also present in statistics and the experimental sciences, with their emphasis on the careful design of experiments to collect the data relevant to a speciﬁc hypothesis. As with quality
43
Chapter 2
Data
issues at the measurement and data collection level, there are many issues that are speciﬁc to particular applications and ﬁelds. Again, we consider only a few of the general issues. Timeliness Some data starts to age as soon as it has been collected. In particular, if the data provides a snapshot of some ongoing phenomenon or process, such as the purchasing behavior of customers or Web browsing patterns, then this snapshot represents reality for only a limited time. If the data is out of date, then so are the models and patterns that are based on it. Relevance The available data must contain the information necessary for the application. Consider the task of building a model that predicts the accident rate for drivers. If information about the age and gender of the driver is omitted, then it is likely that the model will have limited accuracy unless this information is indirectly available through other attributes. Making sure that the objects in a data set are relevant is also challenging. A common problem is sampling bias, which occurs when a sample does not contain diﬀerent types of objects in proportion to their actual occurrence in the population. For example, survey data describes only those who respond to the survey. (Other aspects of sampling are discussed further in Section 2.3.2.) Because the results of a data analysis can reﬂect only the data that is present, sampling bias will typically result in an erroneous analysis. Knowledge about the Data Ideally, data sets are accompanied by documentation that describes diﬀerent aspects of the data; the quality of this documentation can either aid or hinder the subsequent analysis. For example, if the documentation identiﬁes several attributes as being strongly related, these attributes are likely to provide highly redundant information, and we may decide to keep just one. (Consider sales tax and purchase price.) If the documentation is poor, however, and fails to tell us, for example, that the missing values for a particular ﬁeld are indicated with a 9999, then our analysis of the data may be faulty. Other important characteristics are the precision of the data, the type of features (nominal, ordinal, interval, ratio), the scale of measurement (e.g., meters or feet for length), and the origin of the data.
2.3
Data Preprocessing
In this section, we address the issue of which preprocessing steps should be applied to make the data more suitable for data mining. Data preprocessing
44
2.3
Data Preprocessing
is a broad area and consists of a number of diﬀerent strategies and techniques that are interrelated in complex ways. We will present some of the most important ideas and approaches, and try to point out the interrelationships among them. Speciﬁcally, we will discuss the following topics: • • • • • • •
Aggregation Sampling Dimensionality reduction Feature subset selection Feature creation Discretization and binarization Variable transformation
Roughly speaking, these items fall into two categories: selecting data objects and attributes for the analysis or creating/changing the attributes. In both cases the goal is to improve the data mining analysis with respect to time, cost, and quality. Details are provided in the following sections. A quick note on terminology: In the following, we sometimes use synonyms for attribute, such as feature or variable, in order to follow common usage.
2.3.1
Aggregation
Sometimes “less is more” and this is the case with aggregation, the combining of two or more objects into a single object. Consider a data set consisting of transactions (data objects) recording the daily sales of products in various store locations (Minneapolis, Chicago, Paris, . . .) for diﬀerent days over the course of a year. See Table 2.4. One way to aggregate transactions for this data set is to replace all the transactions of a single store with a single storewide transaction. This reduces the hundreds or thousands of transactions that occur daily at a speciﬁc store to a single daily transaction, and the number of data objects is reduced to the number of stores. An obvious issue is how an aggregate transaction is created; i.e., how the values of each attribute are combined across all the records corresponding to a particular location to create the aggregate transaction that represents the sales of a single store or date. Quantitative attributes, such as price, are typically aggregated by taking a sum or an average. A qualitative attribute, such as item, can either be omitted or summarized as the set of all the items that were sold at that location. The data in Table 2.4 can also be viewed as a multidimensional array, where each attribute is a dimension. From this viewpoint, aggregation is the
45
Chapter 2
Data
Table 2.4. Data set containing information about customer purchases. Transaction ID .. . 101123 101123 101124 .. .
Item .. . Watch Battery Shoes .. .
Store Location .. . Chicago Chicago Minneapolis .. .
Date .. . 09/06/04 09/06/04 09/06/04 .. .
Price .. . $25.99 $5.99 $75.00 .. .
... ... ... ...
process of eliminating attributes, such as the type of item, or reducing the number of values for a particular attribute; e.g., reducing the possible values for date from 365 days to 12 months. This type of aggregation is commonly used in Online Analytical Processing (OLAP), which is discussed further in Chapter 3. There are several motivations for aggregation. First, the smaller data sets resulting from data reduction require less memory and processing time, and hence, aggregation may permit the use of more expensive data mining algorithms. Second, aggregation can act as a change of scope or scale by providing a highlevel view of the data instead of a lowlevel view. In the previous example, aggregating over store locations and months gives us a monthly, per store view of the data instead of a daily, per item view. Finally, the behavior of groups of objects or attributes is often more stable than that of individual objects or attributes. This statement reﬂects the statistical fact that aggregate quantities, such as averages or totals, have less variability than the individual objects being aggregated. For totals, the actual amount of variation is larger than that of individual objects (on average), but the percentage of the variation is smaller, while for means, the actual amount of variation is less than that of individual objects (on average). A disadvantage of aggregation is the potential loss of interesting details. In the store example aggregating over months loses information about which day of the week has the highest sales. Example 2.7 (Australian Precipitation). This example is based on precipitation in Australia from the period 1982 to 1993. Figure 2.8(a) shows a histogram for the standard deviation of average monthly precipitation for 3,030 0.5◦ by 0.5◦ grid cells in Australia, while Figure 2.8(b) shows a histogram for the standard deviation of the average yearly precipitation for the same locations. The average yearly precipitation has less variability than the average monthly precipitation. All precipitation measurements (and their standard deviations) are in centimeters.
46
2.3
Data Preprocessing
150
180
140
Number of Land Locations
Number of Land Locations
160
120 100 80 60
100
50
40 20 0
0
2
4
6
8 10 12 Standard Deviation
14
16
(a) Histogram of standard deviation of average monthly precipitation
18
0 0
1
2
3 4 Standard Deviation
5
6
(b) Histogram of standard deviation of average yearly precipitation
Figure 2.8. Histograms of standard deviation for monthly and yearly precipitation in Australia for the period 1982 to 1993.
2.3.2
Sampling
Sampling is a commonly used approach for selecting a subset of the data objects to be analyzed. In statistics, it has long been used for both the preliminary investigation of the data and the ﬁnal data analysis. Sampling can also be very useful in data mining. However, the motivations for sampling in statistics and data mining are often diﬀerent. Statisticians use sampling because obtaining the entire set of data of interest is too expensive or time consuming, while data miners sample because it is too expensive or time consuming to process all the data. In some cases, using a sampling algorithm can reduce the data size to the point where a better, but more expensive algorithm can be used. The key principle for eﬀective sampling is the following: Using a sample will work almost as well as using the entire data set if the sample is representative. In turn, a sample is representative if it has approximately the same property (of interest) as the original set of data. If the mean (average) of the data objects is the property of interest, then a sample is representative if it has a mean that is close to that of the original data. Because sampling is a statistical process, the representativeness of any particular sample will vary, and the best that we can do is choose a sampling scheme that guarantees a high probability of getting a representative sample. As discussed next, this involves choosing the appropriate sample size and sampling techniques.
47
Chapter 2
Data
Sampling Approaches There are many sampling techniques, but only a few of the most basic ones and their variations will be covered here. The simplest type of sampling is simple random sampling. For this type of sampling, there is an equal probability of selecting any particular item. There are two variations on random sampling (and other sampling techniques as well): (1) sampling without replacement—as each item is selected, it is removed from the set of all objects that together constitute the population, and (2) sampling with replacement—objects are not removed from the population as they are selected for the sample. In sampling with replacement, the same object can be picked more than once. The samples produced by the two methods are not much diﬀerent when samples are relatively small compared to the data set size, but sampling with replacement is simpler to analyze since the probability of selecting any object remains constant during the sampling process. When the population consists of diﬀerent types of objects, with widely diﬀerent numbers of objects, simple random sampling can fail to adequately represent those types of objects that are less frequent. This can cause problems when the analysis requires proper representation of all object types. For example, when building classiﬁcation models for rare classes, it is critical that the rare classes be adequately represented in the sample. Hence, a sampling scheme that can accommodate diﬀering frequencies for the items of interest is needed. Stratiﬁed sampling, which starts with prespeciﬁed groups of objects, is such an approach. In the simplest version, equal numbers of objects are drawn from each group even though the groups are of diﬀerent sizes. In another variation, the number of objects drawn from each group is proportional to the size of that group. Example 2.8 (Sampling and Loss of Information). Once a sampling technique has been selected, it is still necessary to choose the sample size. Larger sample sizes increase the probability that a sample will be representative, but they also eliminate much of the advantage of sampling. Conversely, with smaller sample sizes, patterns may be missed or erroneous patterns can be detected. Figure 2.9(a) shows a data set that contains 8000 twodimensional points, while Figures 2.9(b) and 2.9(c) show samples from this data set of size 2000 and 500, respectively. Although most of the structure of this data set is present in the sample of 2000 points, much of the structure is missing in the sample of 500 points.
48
2.3
(a) 8000 points
(b) 2000 points
Data Preprocessing
(c) 500 points
Figure 2.9. Example of the loss of structure with sampling.
Example 2.9 (Determining the Proper Sample Size). To illustrate that determining the proper sample size requires a methodical approach, consider the following task. Given a set of data that consists of a small number of almost equalsized groups, ﬁnd at least one representative point for each of the groups. Assume that the objects in each group are highly similar to each other, but not very similar to objects in diﬀerent groups. Also assume that there are a relatively small number of groups, e.g., 10. Figure 2.10(a) shows an idealized set of clusters (groups) from which these points might be drawn. This problem can be eﬃciently solved using sampling. One approach is to take a small sample of data points, compute the pairwise similarities between points, and then form groups of points that are highly similar. The desired set of representative points is then obtained by taking one point from each of these groups. To follow this approach, however, we need to determine a sample size that would guarantee, with a high probability, the desired outcome; that is, that at least one point will be obtained from each cluster. Figure 2.10(b) shows the probability of getting one object from each of the 10 groups as the sample size runs from 10 to 60. Interestingly, with a sample size of 20, there is little chance (20%) of getting a sample that includes all 10 clusters. Even with a sample size of 30, there is still a moderate chance (almost 40%) of getting a sample that doesn’t contain objects from all 10 clusters. This issue is further explored in the context of clustering by Exercise 4 on page 559.
49
Chapter 2
Data
1
Probability
0.8 0.6 0.4 0.2 0
(a) Ten groups of points.
0
10
20
30 40 Sample Size
50
60
70
(b) Probability a sample contains points from each of 10 groups.
Figure 2.10. Finding representative points from 10 groups.
Progressive Sampling The proper sample size can be diﬃcult to determine, so adaptive or progressive sampling schemes are sometimes used. These approaches start with a small sample, and then increase the sample size until a sample of suﬃcient size has been obtained. While this technique eliminates the need to determine the correct sample size initially, it requires that there be a way to evaluate the sample to judge if it is large enough. Suppose, for instance, that progressive sampling is used to learn a predictive model. Although the accuracy of predictive models increases as the sample size increases, at some point the increase in accuracy levels oﬀ. We want to stop increasing the sample size at this levelingoﬀ point. By keeping track of the change in accuracy of the model as we take progressively larger samples, and by taking other samples close to the size of the current one, we can get an estimate as to how close we are to this levelingoﬀ point, and thus, stop sampling.
2.3.3
Dimensionality Reduction
Data sets can have a large number of features. Consider a set of documents, where each document is represented by a vector whose components are the frequencies with which each word occurs in the document. In such cases,
50
2.3
Data Preprocessing
there are typically thousands or tens of thousands of attributes (components), one for each word in the vocabulary. As another example, consider a set of time series consisting of the daily closing price of various stocks over a period of 30 years. In this case, the attributes, which are the prices on speciﬁc days, again number in the thousands. There are a variety of beneﬁts to dimensionality reduction. A key beneﬁt is that many data mining algorithms work better if the dimensionality—the number of attributes in the data—is lower. This is partly because dimensionality reduction can eliminate irrelevant features and reduce noise and partly because of the curse of dimensionality, which is explained below. Another beneﬁt is that a reduction of dimensionality can lead to a more understandable model because the model may involve fewer attributes. Also, dimensionality reduction may allow the data to be more easily visualized. Even if dimensionality reduction doesn’t reduce the data to two or three dimensions, data is often visualized by looking at pairs or triplets of attributes, and the number of such combinations is greatly reduced. Finally, the amount of time and memory required by the data mining algorithm is reduced with a reduction in dimensionality. The term dimensionality reduction is often reserved for those techniques that reduce the dimensionality of a data set by creating new attributes that are a combination of the old attributes. The reduction of dimensionality by selecting new attributes that are a subset of the old is known as feature subset selection or feature selection. It will be discussed in Section 2.3.4. In the remainder of this section, we brieﬂy introduce two important topics: the curse of dimensionality and dimensionality reduction techniques based on linear algebra approaches such as principal components analysis (PCA). More details on dimensionality reduction can be found in Appendix B. The Curse of Dimensionality The curse of dimensionality refers to the phenomenon that many types of data analysis become signiﬁcantly harder as the dimensionality of the data increases. Speciﬁcally, as dimensionality increases, the data becomes increasingly sparse in the space that it occupies. For classiﬁcation, this can mean that there are not enough data objects to allow the creation of a model that reliably assigns a class to all possible objects. For clustering, the deﬁnitions of density and the distance between points, which are critical for clustering, become less meaningful. (This is discussed further in Sections 9.1.2, 9.4.5, and 9.4.7.) As a result, many clustering and classiﬁcation algorithms (and other
51
Chapter 2
Data
data analysis algorithms) have trouble with highdimensional data—reduced classiﬁcation accuracy and poor quality clusters. Linear Algebra Techniques for Dimensionality Reduction Some of the most common approaches for dimensionality reduction, particularly for continuous data, use techniques from linear algebra to project the data from a highdimensional space into a lowerdimensional space. Principal Components Analysis (PCA) is a linear algebra technique for continuous attributes that ﬁnds new attributes (principal components) that (1) are linear combinations of the original attributes, (2) are orthogonal (perpendicular) to each other, and (3) capture the maximum amount of variation in the data. For example, the ﬁrst two principal components capture as much of the variation in the data as is possible with two orthogonal attributes that are linear combinations of the original attributes. Singular Value Decomposition (SVD) is a linear algebra technique that is related to PCA and is also commonly used for dimensionality reduction. For additional details, see Appendices A and B.
2.3.4
Feature Subset Selection
Another way to reduce the dimensionality is to use only a subset of the features. While it might seem that such an approach would lose information, this is not the case if redundant and irrelevant features are present. Redundant features duplicate much or all of the information contained in one or more other attributes. For example, the purchase price of a product and the amount of sales tax paid contain much of the same information. Irrelevant features contain almost no useful information for the data mining task at hand. For instance, students’ ID numbers are irrelevant to the task of predicting students’ grade point averages. Redundant and irrelevant features can reduce classiﬁcation accuracy and the quality of the clusters that are found. While some irrelevant and redundant attributes can be eliminated immediately by using common sense or domain knowledge, selecting the best subset of features frequently requires a systematic approach. The ideal approach to feature selection is to try all possible subsets of features as input to the data mining algorithm of interest, and then take the subset that produces the best results. This method has the advantage of reﬂecting the objective and bias of the data mining algorithm that will eventually be used. Unfortunately, since the number of subsets involving n attributes is 2n , such an approach is impractical in most situations and alternative strategies are needed. There are three standard approaches to feature selection: embedded, ﬁlter, and wrapper.
52
2.3
Data Preprocessing
Embedded approaches Feature selection occurs naturally as part of the data mining algorithm. Speciﬁcally, during the operation of the data mining algorithm, the algorithm itself decides which attributes to use and which to ignore. Algorithms for building decision tree classiﬁers, which are discussed in Chapter 4, often operate in this manner. Filter approaches Features are selected before the data mining algorithm is run, using some approach that is independent of the data mining task. For example, we might select sets of attributes whose pairwise correlation is as low as possible. Wrapper approaches These methods use the target data mining algorithm as a black box to ﬁnd the best subset of attributes, in a way similar to that of the ideal algorithm described above, but typically without enumerating all possible subsets. Since the embedded approaches are algorithmspeciﬁc, only the ﬁlter and wrapper approaches will be discussed further here. An Architecture for Feature Subset Selection It is possible to encompass both the ﬁlter and wrapper approaches within a common architecture. The feature selection process is viewed as consisting of four parts: a measure for evaluating a subset, a search strategy that controls the generation of a new subset of features, a stopping criterion, and a validation procedure. Filter methods and wrapper methods diﬀer only in the way in which they evaluate a subset of features. For a wrapper method, subset evaluation uses the target data mining algorithm, while for a ﬁlter approach, the evaluation technique is distinct from the target data mining algorithm. The following discussion provides some details of this approach, which is summarized in Figure 2.11. Conceptually, feature subset selection is a search over all possible subsets of features. Many diﬀerent types of search strategies can be used, but the search strategy should be computationally inexpensive and should ﬁnd optimal or near optimal sets of features. It is usually not possible to satisfy both requirements, and thus, tradeoﬀs are necessary. An integral part of the search is an evaluation step to judge how the current subset of features compares to others that have been considered. This requires an evaluation measure that attempts to determine the goodness of a subset of attributes with respect to a particular data mining task, such as classiﬁcation
53
Chapter 2
Selected Attributes
Data Done
Stopping Criterion
Evaluation
Not Done Validation Procedure
Attributes
Search Strategy
Subset of Attributes
Figure 2.11. Flowchart of a feature subset selection process.
or clustering. For the ﬁlter approach, such measures attempt to predict how well the actual data mining algorithm will perform on a given set of attributes. For the wrapper approach, where evaluation consists of actually running the target data mining application, the subset evaluation function is simply the criterion normally used to measure the result of the data mining. Because the number of subsets can be enormous and it is impractical to examine them all, some sort of stopping criterion is necessary. This strategy is usually based on one or more conditions involving the following: the number of iterations, whether the value of the subset evaluation measure is optimal or exceeds a certain threshold, whether a subset of a certain size has been obtained, whether simultaneous size and evaluation criteria have been achieved, and whether any improvement can be achieved by the options available to the search strategy. Finally, once a subset of features has been selected, the results of the target data mining algorithm on the selected subset should be validated. A straightforward evaluation approach is to run the algorithm with the full set of features and compare the full results to results obtained using the subset of features. Hopefully, the subset of features will produce results that are better than or almost as good as those produced when using all features. Another validation approach is to use a number of diﬀerent feature selection algorithms to obtain subsets of features and then compare the results of running the data mining algorithm on each subset.
54
2.3
Data Preprocessing
Feature Weighting Feature weighting is an alternative to keeping or eliminating features. More important features are assigned a higher weight, while less important features are given a lower weight. These weights are sometimes assigned based on domain knowledge about the relative importance of features. Alternatively, they may be determined automatically. For example, some classiﬁcation schemes, such as support vector machines (Chapter 5), produce classiﬁcation models in which each feature is given a weight. Features with larger weights play a more important role in the model. The normalization of objects that takes place when computing the cosine similarity (Section 2.4.5) can also be regarded as a type of feature weighting.
2.3.5
Feature Creation
It is frequently possible to create, from the original attributes, a new set of attributes that captures the important information in a data set much more eﬀectively. Furthermore, the number of new attributes can be smaller than the number of original attributes, allowing us to reap all the previously described beneﬁts of dimensionality reduction. Three related methodologies for creating new attributes are described next: feature extraction, mapping the data to a new space, and feature construction. Feature Extraction The creation of a new set of features from the original raw data is known as feature extraction. Consider a set of photographs, where each photograph is to be classiﬁed according to whether or not it contains a human face. The raw data is a set of pixels, and as such, is not suitable for many types of classiﬁcation algorithms. However, if the data is processed to provide higherlevel features, such as the presence or absence of certain types of edges and areas that are highly correlated with the presence of human faces, then a much broader set of classiﬁcation techniques can be applied to this problem. Unfortunately, in the sense in which it is most commonly used, feature extraction is highly domainspeciﬁc. For a particular ﬁeld, such as image processing, various features and the techniques to extract them have been developed over a period of time, and often these techniques have limited applicability to other ﬁelds. Consequently, whenever data mining is applied to a relatively new area, a key task is the development of new features and feature extraction methods.
55
Chapter 2
Data
1
15
300
10
250
5
200
0
150
5
100
10
50
0.5
0
0.5
1
0
0.2
0.4 0.6 Time (seconds)
0.8
(a) Two time series.
1
15 0
0.2
0.4 0.6 Time (seconds)
0.8
(b) Noisy time series.
1
0 0
10
20
30
40 50 60 Frequency
70
80
90
(c) Power spectrum
Figure 2.12. Application of the Fourier transform to identify the underlying frequencies in time series data.
Mapping the Data to a New Space A totally diﬀerent view of the data can reveal important and interesting features. Consider, for example, time series data, which often contains periodic patterns. If there is only a single periodic pattern and not much noise, then the pattern is easily detected. If, on the other hand, there are a number of periodic patterns and a signiﬁcant amount of noise is present, then these patterns are hard to detect. Such patterns can, nonetheless, often be detected by applying a Fourier transform to the time series in order to change to a representation in which frequency information is explicit. In the example that follows, it will not be necessary to know the details of the Fourier transform. It is enough to know that, for each time series, the Fourier transform produces a new data object whose attributes are related to frequencies. Example 2.10 (Fourier Analysis). The time series presented in Figure 2.12(b) is the sum of three other time series, two of which are shown in Figure 2.12(a) and have frequencies of 7 and 17 cycles per second, respectively. The third time series is random noise. Figure 2.12(c) shows the power spectrum that can be computed after applying a Fourier transform to the original time series. (Informally, the power spectrum is proportional to the square of each frequency attribute.) In spite of the noise, there are two peaks that correspond to the periods of the two original, nonnoisy time series. Again, the main point is that better features can reveal important aspects of the data.
56
2.3
Data Preprocessing
Many other sorts of transformations are also possible. Besides the Fourier transform, the wavelet transform has also proven very useful for time series and other types of data. Feature Construction Sometimes the features in the original data sets have the necessary information, but it is not in a form suitable for the data mining algorithm. In this situation, one or more new features constructed out of the original features can be more useful than the original features. Example 2.11 (Density). To illustrate this, consider a data set consisting of information about historical artifacts, which, along with other information, contains the volume and mass of each artifact. For simplicity, assume that these artifacts are made of a small number of materials (wood, clay, bronze, gold) and that we want to classify the artifacts with respect to the material of which they are made. In this case, a density feature constructed from the mass and volume features, i.e., density = mass/volume, would most directly yield an accurate classiﬁcation. Although there have been some attempts to automatically perform feature construction by exploring simple mathematical combinations of existing attributes, the most common approach is to construct features using domain expertise.
2.3.6
Discretization and Binarization
Some data mining algorithms, especially certain classiﬁcation algorithms, require that the data be in the form of categorical attributes. Algorithms that ﬁnd association patterns require that the data be in the form of binary attributes. Thus, it is often necessary to transform a continuous attribute into a categorical attribute (discretization), and both continuous and discrete attributes may need to be transformed into one or more binary attributes (binarization). Additionally, if a categorical attribute has a large number of values (categories), or some values occur infrequently, then it may be beneﬁcial for certain data mining tasks to reduce the number of categories by combining some of the values. As with feature selection, the best discretization and binarization approach is the one that “produces the best result for the data mining algorithm that will be used to analyze the data.” It is typically not practical to apply such a criterion directly. Consequently, discretization or binarization is performed in
57
Chapter 2
Data
Table 2.5. Conversion of a categorical attribute to three binary attributes. Categorical Value awful poor OK good great
Integer Value 0 1 2 3 4
x1 0 0 0 0 1
x2 0 0 1 1 0
x3 0 1 0 1 0
Table 2.6. Conversion of a categorical attribute to five asymmetric binary attributes. Categorical Value awful poor OK good great
Integer Value 0 1 2 3 4
x1 1 0 0 0 0
x2 0 1 0 0 0
x3 0 0 1 0 0
x4 0 0 0 1 0
x5 0 0 0 0 1
a way that satisﬁes a criterion that is thought to have a relationship to good performance for the data mining task being considered. Binarization A simple technique to binarize a categorical attribute is the following: If there are m categorical values, then uniquely assign each original value to an integer in the interval [0, m − 1]. If the attribute is ordinal, then order must be maintained by the assignment. (Note that even if the attribute is originally represented using integers, this process is necessary if the integers are not in the interval [0, m − 1].) Next, convert each of these m integers to a binary number. Since n = log2 (m) binary digits are required to represent these integers, represent these binary numbers using n binary attributes. To illustrate, a categorical variable with 5 values {awful, poor, OK, good, great} would require three binary variables x1 , x2 , and x3 . The conversion is shown in Table 2.5. Such a transformation can cause complications, such as creating unintended relationships among the transformed attributes. For example, in Table 2.5, attributes x2 and x3 are correlated because information about the good value is encoded using both attributes. Furthermore, association analysis requires asymmetric binary attributes, where only the presence of the attribute (value = 1) is important. For association problems, it is therefore necessary to introduce one binary attribute for each categorical value, as in Table 2.6. If the
58
2.3
Data Preprocessing
number of resulting attributes is too large, then the techniques described below can be used to reduce the number of categorical values before binarization. Likewise, for association problems, it may be necessary to replace a single binary attribute with two asymmetric binary attributes. Consider a binary attribute that records a person’s gender, male or female. For traditional association rule algorithms, this information needs to be transformed into two asymmetric binary attributes, one that is a 1 only when the person is male and one that is a 1 only when the person is female. (For asymmetric binary attributes, the information representation is somewhat ineﬃcient in that two bits of storage are required to represent each bit of information.) Discretization of Continuous Attributes Discretization is typically applied to attributes that are used in classiﬁcation or association analysis. In general, the best discretization depends on the algorithm being used, as well as the other attributes being considered. Typically, however, the discretization of an attribute is considered in isolation. Transformation of a continuous attribute to a categorical attribute involves two subtasks: deciding how many categories to have and determining how to map the values of the continuous attribute to these categories. In the ﬁrst step, after the values of the continuous attribute are sorted, they are then divided into n intervals by specifying n − 1 split points. In the second, rather trivial step, all the values in one interval are mapped to the same categorical value. Therefore, the problem of discretization is one of deciding how many split points to choose and where to place them. The result can be represented either as a set of intervals {(x0 , x1 ], (x1 , x2 ], . . . , (xn−1 , xn )}, where x0 and xn may be +∞ or −∞, respectively, or equivalently, as a series of inequalities x0 < x ≤ x1 , . . . , xn−1 < x < xn . Unsupervised Discretization A basic distinction between discretization methods for classiﬁcation is whether class information is used (supervised) or not (unsupervised). If class information is not used, then relatively simple approaches are common. For instance, the equal width approach divides the range of the attribute into a userspeciﬁed number of intervals each having the same width. Such an approach can be badly aﬀected by outliers, and for that reason, an equal frequency (equal depth) approach, which tries to put the same number of objects into each interval, is often preferred. As another example of unsupervised discretization, a clustering method, such as Kmeans (see Chapter 8), can also be used. Finally, visually inspecting the data can sometimes be an eﬀective approach.
59
Chapter 2
Data
Example 2.12 (Discretization Techniques). This example demonstrates how these approaches work on an actual data set. Figure 2.13(a) shows data points belonging to four diﬀerent groups, along with two outliers—the large dots on either end. The techniques of the previous paragraph were applied to discretize the x values of these data points into four categorical values. (Points in the data set have a random y component to make it easy to see how many points are in each group.) Visually inspecting the data works quite well, but is not automatic, and thus, we focus on the other three approaches. The split points produced by the techniques equal width, equal frequency, and Kmeans are shown in Figures 2.13(b), 2.13(c), and 2.13(d), respectively. The split points are represented as dashed lines. If we measure the performance of a discretization technique by the extent to which diﬀerent objects in diﬀerent groups are assigned the same categorical value, then Kmeans performs best, followed by equal frequency, and ﬁnally, equal width. Supervised Discretization The discretization methods described above are usually better than no discretization, but keeping the end purpose in mind and using additional information (class labels) often produces better results. This should not be surprising, since an interval constructed with no knowledge of class labels often contains a mixture of class labels. A conceptually simple approach is to place the splits in a way that maximizes the purity of the intervals. In practice, however, such an approach requires potentially arbitrary decisions about the purity of an interval and the minimum size of an interval. To overcome such concerns, some statistically based approaches start with each attribute value as a separate interval and create larger intervals by merging adjacent intervals that are similar according to a statistical test. Entropybased approaches are one of the most promising approaches to discretization, and a simple approach based on entropy will be presented. First, it is necessary to deﬁne entropy. Let k be the number of diﬀerent class labels, mi be the number of values in the ith interval of a partition, and mij be the number of values of class j in interval i. Then the entropy ei of the ith interval is given by the equation ei =
k
pij log2 pij ,
i=1
where pij = mij /mi is the probability (fraction of values) of class j in the ith interval. The total entropy, e, of the partition is the weighted average of the individual interval entropies, i.e.,
60
2.3
0
5
10
15
20
0
(a) Original data.
0
5
10
15
Data Preprocessing
5
10
15
20
(b) Equal width discretization.
20
0
(c) Equal frequency discretization.
5
10
15
20
(d) Kmeans discretization.
Figure 2.13. Different discretization techniques.
e=
n
wi ei ,
i=1
where m is the number of values, wi = mi /m is the fraction of values in the ith interval, and n is the number of intervals. Intuitively, the entropy of an interval is a measure of the purity of an interval. If an interval contains only values of one class (is perfectly pure), then the entropy is 0 and it contributes
61
Chapter 2
Data
nothing to the overall entropy. If the classes of values in an interval occur equally often (the interval is as impure as possible), then the entropy is a maximum. A simple approach for partitioning a continuous attribute starts by bisecting the initial values so that the resulting two intervals give minimum entropy. This technique only needs to consider each value as a possible split point, because it is assumed that intervals contain ordered sets of values. The splitting process is then repeated with another interval, typically choosing the interval with the worst (highest) entropy, until a userspeciﬁed number of intervals is reached, or a stopping criterion is satisﬁed. Example 2.13 (Discretization of Two Attributes). This method was used to independently discretize both the x and y attributes of the twodimensional data shown in Figure 2.14. In the ﬁrst discretization, shown in Figure 2.14(a), the x and y attributes were both split into three intervals. (The dashed lines indicate the split points.) In the second discretization, shown in Figure 2.14(b), the x and y attributes were both split into ﬁve intervals. This simple example illustrates two aspects of discretization. First, in two dimensions, the classes of points are well separated, but in one dimension, this is not so. In general, discretizing each attribute separately often guarantees suboptimal results. Second, ﬁve intervals work better than three, but six intervals do not improve the discretization much, at least in terms of entropy. (Entropy values and results for six intervals are not shown.) Consequently, it is desirable to have a stopping criterion that automatically ﬁnds the right number of partitions. Categorical Attributes with Too Many Values Categorical attributes can sometimes have too many values. If the categorical attribute is an ordinal attribute, then techniques similar to those for continuous attributes can be used to reduce the number of categories. If the categorical attribute is nominal, however, then other approaches are needed. Consider a university that has a large number of departments. Consequently, a department name attribute might have dozens of diﬀerent values. In this situation, we could use our knowledge of the relationships among diﬀerent departments to combine departments into larger groups, such as engineering, social sciences, or biological sciences. If domain knowledge does not serve as a useful guide or such an approach results in poor classiﬁcation performance, then it is necessary to use a more empirical approach, such as grouping values
62
2.3 5
4
4
3
3
y
y
5
Data Preprocessing
2
2
1
1
0 0
1
2
3
4
5
x
(a) Three intervals
0 0
1
2
x
3
4
5
(b) Five intervals
Figure 2.14. Discretizing x and y attributes for four groups (classes) of points.
together only if such a grouping results in improved classiﬁcation accuracy or achieves some other data mining objective.
2.3.7
Variable Transformation
A variable transformation refers to a transformation that is applied to all the values of a variable. (We use the term variable instead of attribute to adhere to common usage, although we will also refer to attribute transformation on occasion.) In other words, for each object, the transformation is applied to the value of the variable for that object. For example, if only the magnitude of a variable is important, then the values of the variable can be transformed by taking the absolute value. In the following section, we discuss two important types of variable transformations: simple functional transformations and normalization. Simple Functions For this type of variable transformation, a simple mathematical function is applied to each value individually. If x is a variable, then examples of such √ transformations include xk , log x, ex , x, 1/x, sin x, or x. In statistics, variable transformations, especially sqrt, log, and 1/x, are often used to transform data that does not have a Gaussian (normal) distribution into data that does. While this can be important, other reasons often take precedence in data min
63
Chapter 2
Data
ing. Suppose the variable of interest is the number of data bytes in a session, and the number of bytes ranges from 1 to 1 billion. This is a huge range, and it may be advantageous to compress it by using a log10 transformation. In this case, sessions that transferred 108 and 109 bytes would be more similar to each other than sessions that transferred 10 and 1000 bytes (9 − 8 = 1 versus 3 − 1 = 2). For some applications, such as network intrusion detection, this may be what is desired, since the ﬁrst two sessions most likely represent transfers of large ﬁles, while the latter two sessions could be two quite distinct types of sessions. Variable transformations should be applied with caution since they change the nature of the data. While this is what is desired, there can be problems if the nature of the transformation is not fully appreciated. For instance, the transformation 1/x reduces the magnitude of values that are 1 or larger, but increases the magnitude of values between 0 and 1. To illustrate, the values {1, 2, 3} go to {1, 12 , 13 }, but the values {1, 12 , 13 } go to {1, 2, 3}. Thus, for all sets of values, the transformation 1/x reverses the order. To help clarify the eﬀect of a transformation, it is important to ask questions such as the following: Does the order need to be maintained? Does the transformation apply to all values, especially negative values and 0? What is the eﬀect of the transformation on the values between 0 and 1? Exercise 17 on page 92 explores other aspects of variable transformation. Normalization or Standardization Another common type of variable transformation is the standardization or normalization of a variable. (In the data mining community the terms are often used interchangeably. In statistics, however, the term normalization can be confused with the transformations used for making a variable normal, i.e., Gaussian.) The goal of standardization or normalization is to make an entire set of values have a particular property. A traditional example is that of “standardizing a variable” in statistics. If x is the mean (average) of the attribute values and sx is their standard deviation, then the transformation x = (x − x)/sx creates a new variable that has a mean of 0 and a standard deviation of 1. If diﬀerent variables are to be combined in some way, then such a transformation is often necessary to avoid having a variable with large values dominate the results of the calculation. To illustrate, consider comparing people based on two variables: age and income. For any two people, the diﬀerence in income will likely be much higher in absolute terms (hundreds or thousands of dollars) than the diﬀerence in age (less than 150). If the diﬀerences in the range of values of age and income are not taken into account, then
64
2.4
Measures of Similarity and Dissimilarity
the comparison between people will be dominated by diﬀerences in income. In particular, if the similarity or dissimilarity of two people is calculated using the similarity or dissimilarity measures deﬁned later in this chapter, then in many cases, such as that of Euclidean distance, the income values will dominate the calculation. The mean and standard deviation are strongly aﬀected by outliers, so the above transformation is often modiﬁed. First, the mean is replaced by the median, i.e., the middle value. Second, the standard deviation is replaced by the absolute standard deviation. Speciﬁcally, ifx is a variable, then the absolute standard deviation of x is given by σA = m i=1 xi − µ, where xi is th the i value of the variable, m is the number of objects, and µ is either the mean or median. Other approaches for computing estimates of the location (center) and spread of a set of values in the presence of outliers are described in Sections 3.2.3 and 3.2.4, respectively. These measures can also be used to deﬁne a standardization transformation.
2.4
Measures of Similarity and Dissimilarity
Similarity and dissimilarity are important because they are used by a number of data mining techniques, such as clustering, nearest neighbor classiﬁcation, and anomaly detection. In many cases, the initial data set is not needed once these similarities or dissimilarities have been computed. Such approaches can be viewed as transforming the data to a similarity (dissimilarity) space and then performing the analysis. We begin with a discussion of the basics: highlevel deﬁnitions of similarity and dissimilarity, and a discussion of how they are related. For convenience, the term proximity is used to refer to either similarity or dissimilarity. Since the proximity between two objects is a function of the proximity between the corresponding attributes of the two objects, we ﬁrst describe how to measure the proximity between objects having only one simple attribute, and then consider proximity measures for objects with multiple attributes. This includes measures such as correlation and Euclidean distance, which are useful for dense data such as time series or twodimensional points, as well as the Jaccard and cosine similarity measures, which are useful for sparse data like documents. Next, we consider several important issues concerning proximity measures. The section concludes with a brief discussion of how to select the right proximity measure.
65
Chapter 2
2.4.1
Data
Basics
Deﬁnitions Informally, the similarity between two objects is a numerical measure of the degree to which the two objects are alike. Consequently, similarities are higher for pairs of objects that are more alike. Similarities are usually nonnegative and are often between 0 (no similarity) and 1 (complete similarity). The dissimilarity between two objects is a numerical measure of the degree to which the two objects are diﬀerent. Dissimilarities are lower for more similar pairs of objects. Frequently, the term distance is used as a synonym for dissimilarity, although, as we shall see, distance is often used to refer to a special class of dissimilarities. Dissimilarities sometimes fall in the interval [0, 1], but it is also common for them to range from 0 to ∞. Transformations Transformations are often applied to convert a similarity to a dissimilarity, or vice versa, or to transform a proximity measure to fall within a particular range, such as [0,1]. For instance, we may have similarities that range from 1 to 10, but the particular algorithm or software package that we want to use may be designed to only work with dissimilarities, or it may only work with similarities in the interval [0,1]. We discuss these issues here because we will employ such transformations later in our discussion of proximity. In addition, these issues are relatively independent of the details of speciﬁc proximity measures. Frequently, proximity measures, especially similarities, are deﬁned or transformed to have values in the interval [0,1]. Informally, the motivation for this is to use a scale in which a proximity value indicates the fraction of similarity (or dissimilarity) between two objects. Such a transformation is often relatively straightforward. For example, if the similarities between objects range from 1 (not at all similar) to 10 (completely similar), we can make them fall within the range [0, 1] by using the transformation s = (s − 1)/9, where s and s are the original and new similarity values, respectively. In the more general case, the transformation of similarities to the interval [0, 1] is given by the expression s = (s − min s)/(max s − min s), where max s and min s are the maximum and minimum similarity values, respectively. Likewise, dissimilarity measures with a ﬁnite range can be mapped to the interval [0,1] by using the formula d = (d − min d)/(max d − min d). There can be various complications in mapping proximity measures to the interval [0, 1], however. If, for example, the proximity measure originally takes
66
2.4
Measures of Similarity and Dissimilarity
values in the interval [0,∞], then a nonlinear transformation is needed and values will not have the same relationship to one another on the new scale. Consider the transformation d = d/(1 + d) for a dissimilarity measure that ranges from 0 to ∞. The dissimilarities 0, 0.5, 2, 10, 100, and 1000 will be transformed into the new dissimilarities 0, 0.33, 0.67, 0.90, 0.99, and 0.999, respectively. Larger values on the original dissimilarity scale are compressed into the range of values near 1, but whether or not this is desirable depends on the application. Another complication is that the meaning of the proximity measure may be changed. For example, correlation, which is discussed later, is a measure of similarity that takes values in the interval [1,1]. Mapping these values to the interval [0,1] by taking the absolute value loses information about the sign, which can be important in some applications. See Exercise 22 on page 94. Transforming similarities to dissimilarities and vice versa is also relatively straightforward, although we again face the issues of preserving meaning and changing a linear scale into a nonlinear scale. If the similarity (or dissimilarity) falls in the interval [0,1], then the dissimilarity can be deﬁned as d = 1 − s (s = 1 − d). Another simple approach is to deﬁne similarity as the negative of the dissimilarity (or vice versa). To illustrate, the dissimilarities 0, 1, 10, and 100 can be transformed into the similarities 0, −1, −10, and −100, respectively. The similarities resulting from the negation transformation are not restricted to the range [0, 1], but if that is desired, then transformations such as 1 d−min d , s = e−d , or s = 1 − max s = d+1 d−min d can be used. For the transformation 1 s = d+1 , the dissimilarities 0, 1, 10, 100 are transformed into 1, 0.5, 0.09, 0.01, respectively. For s = e−d , they become 1.00, 0.37, 0.00, 0.00, respectively, d−min d while for s = 1 − max d−min d they become 1.00, 0.99, 0.00, 0.00, respectively. In this discussion, we have focused on converting dissimilarities to similarities. Conversion in the opposite direction is considered in Exercise 23 on page 94. In general, any monotonic decreasing function can be used to convert dissimilarities to similarities, or vice versa. Of course, other factors also must be considered when transforming similarities to dissimilarities, or vice versa, or when transforming the values of a proximity measure to a new scale. We have mentioned issues related to preserving meaning, distortion of scale, and requirements of data analysis tools, but this list is certainly not exhaustive.
2.4.2
Similarity and Dissimilarity between Simple Attributes
The proximity of objects with a number of attributes is typically deﬁned by combining the proximities of individual attributes, and thus, we ﬁrst discuss
67
Chapter 2
Data
proximity between objects having a single attribute. Consider objects described by one nominal attribute. What would it mean for two such objects to be similar? Since nominal attributes only convey information about the distinctness of objects, all we can say is that two objects either have the same value or they do not. Hence, in this case similarity is traditionally deﬁned as 1 if attribute values match, and as 0 otherwise. A dissimilarity would be deﬁned in the opposite way: 0 if the attribute values match, and 1 if they do not. For objects with a single ordinal attribute, the situation is more complicated because information about order should be taken into account. Consider an attribute that measures the quality of a product, e.g., a candy bar, on the scale {poor, fair, OK, good, wonderful }. It would seem reasonable that a product, P1, which is rated wonderful, would be closer to a product P2, which is rated good, than it would be to a product P3, which is rated OK. To make this observation quantitative, the values of the ordinal attribute are often mapped to successive integers, beginning at 0 or 1, e.g., {poor =0, fair =1, OK =2, good=3, wonderful =4}. Then, d(P1, P2) = 3 − 2 = 1 or, if we want the dissimilarity to fall between 0 and 1, d(P1, P2) = 3−2 4 = 0.25. A similarity for ordinal attributes can then be deﬁned as s = 1 − d. This deﬁnition of similarity (dissimilarity) for an ordinal attribute should make the reader a bit uneasy since this assumes equal intervals, and this is not so. Otherwise, we would have an interval or ratio attribute. Is the diﬀerence between the values fair and good really the same as that between the values OK and wonderful ? Probably not, but in practice, our options are limited, and in the absence of more information, this is the standard approach for deﬁning proximity between ordinal attributes. For interval or ratio attributes, the natural measure of dissimilarity between two objects is the absolute diﬀerence of their values. For example, we might compare our current weight and our weight a year ago by saying “I am ten pounds heavier.” In cases such as these, the dissimilarities typically range from 0 to ∞, rather than from 0 to 1. The similarity of interval or ratio attributes is typically expressed by transforming a similarity into a dissimilarity, as previously described. Table 2.7 summarizes this discussion. In this table, x and y are two objects that have one attribute of the indicated type. Also, d(x, y) and s(x, y) are the dissimilarity and similarity between x and y, respectively. Other approaches are possible; these are the most common ones. The following two sections consider more complicated measures of proximity between objects that involve multiple attributes: (1) dissimilarities between data objects and (2) similarities between data objects. This division
68
2.4
Measures of Similarity and Dissimilarity
Table 2.7. Similarity and dissimilarity for simple attributes Attribute Type Nominal Ordinal Interval or Ratio
Dissimilarity
Similarity
0 if x = y d= 1 if x = y d = x − y/(n − 1) (values mapped to integers 0 to n−1, where n is the number of values) d = x − y
s=
1 0
if x = y if x = y
s=1−d 1 s = −d, s = 1+d , s = e−d , d−min d s = 1 − max d−min d
allows us to more naturally display the underlying motivations for employing various proximity measures. We emphasize, however, that similarities can be transformed into dissimilarities and vice versa using the approaches described earlier.
2.4.3
Dissimilarities between Data Objects
In this section, we discuss various kinds of dissimilarities. We begin with a discussion of distances, which are dissimilarities with certain properties, and then provide examples of more general kinds of dissimilarities. Distances We ﬁrst present some examples, and then oﬀer a more formal description of distances in terms of the properties common to all distances. The Euclidean distance, d, between two points, x and y, in one, two, three, or higherdimensional space, is given by the following familiar formula: n d(x, y) = (xk − yk )2 ,
(2.1)
k=1
where n is the number of dimensions and xk and yk are, respectively, the k th attributes (components) of x and y. We illustrate this formula with Figure 2.15 and Tables 2.8 and 2.9, which show a set of points, the x and y coordinates of these points, and the distance matrix containing the pairwise distances of these points.
69
Chapter 2
Data
The Euclidean distance measure given in Equation 2.1 is generalized by the Minkowski distance metric shown in Equation 2.2, d(x, y) =
n
1/r xk − yk r
,
(2.2)
k=1
where r is a parameter. The following are the three most common examples of Minkowski distances. • r = 1. City block (Manhattan, taxicab, L1 norm) distance. A common example is the Hamming distance, which is the number of bits that are diﬀerent between two objects that have only binary attributes, i.e., between two binary vectors. • r = 2. Euclidean distance (L2 norm). • r = ∞. Supremum (Lmax or L∞ norm) distance. This is the maximum diﬀerence between any attribute of the objects. More formally, the L∞ distance is deﬁned by Equation 2.3 d(x, y) = lim
r→∞
n
1/r xk − yk r
.
(2.3)
k=1
The r parameter should not be confused with the number of dimensions (attributes) n. The Euclidean, Manhattan, and supremum distances are deﬁned for all values of n: 1, 2, 3, . . ., and specify diﬀerent ways of combining the diﬀerences in each dimension (attribute) into an overall distance. Tables 2.10 and 2.11, respectively, give the proximity matrices for the L1 and L∞ distances using data from Table 2.8. Notice that all these distance matrices are symmetric; i.e., the ij th entry is the same as the jith entry. In Table 2.9, for instance, the fourth row of the ﬁrst column and the fourth column of the ﬁrst row both contain the value 5.1. Distances, such as the Euclidean distance, have some wellknown properties. If d(x, y) is the distance between two points, x and y, then the following properties hold. 1. Positivity (a) d(x, x) ≥ 0 for all x and y, (b) d(x, y) = 0 only if x = y.
70
2.4
Measures of Similarity and Dissimilarity
3 2
p1
y
p3
p4
1 p2 0 1
2
3 x
4
5
6
Figure 2.15. Four twodimensional points. Table 2.8. x and y coordinates of four points. point p1 p2 p3 p4
x coordinate 0 2 3 5
y coordinate 2 0 1 1
Table 2.10. L1 distance matrix for Table 2.8. L1 p1 p2 p3 p4
p1 0.0 4.0 4.0 6.0
p2 4.0 0.0 2.0 4.0
p3 4.0 2.0 0.0 2.0
p4 6.0 4.0 2.0 0.0
Table 2.9. Euclidean distance matrix for Table 2.8. p1 p2 p3 p4 p1 0.0 2.8 3.2 5.1 p2 2.8 0.0 1.4 3.2 p3 3.2 1.4 0.0 2.0 p4 5.1 3.2 2.0 0.0 Table 2.11. L∞ distance matrix for Table 2.8. L∞ p1 p2 p3 p4
p1 0.0 2.0 3.0 5.0
p2 2.0 0.0 1.0 3.0
p3 3.0 1.0 0.0 2.0
p4 5.0 3.0 2.0 0.0
2. Symmetry d(x, y) = d(y, x) for all x and y. 3. Triangle Inequality d(x, z) ≤ d(x, y) + d(y, z) for all points x, y, and z. Measures that satisfy all three properties are known as metrics. Some people only use the term distance for dissimilarity measures that satisfy these properties, but that practice is often violated. The three properties described here are useful, as well as mathematically pleasing. Also, if the triangle inequality holds, then this property can be used to increase the eﬃciency of techniques (including clustering) that depend on distances possessing this property. (See Exercise 25.) Nonetheless, many dissimilarities do not satisfy one or more of the metric properties. We give two examples of such measures.
71
Chapter 2
Data
Example 2.14 (Nonmetric Dissimilarities: Set Diﬀerences). This example is based on the notion of the diﬀerence of two sets, as deﬁned in set theory. Given two sets A and B, A − B is the set of elements of A that are not in B. For example, if A = {1, 2, 3, 4} and B = {2, 3, 4}, then A − B = {1} and B − A = ∅, the empty set. We can deﬁne the distance d between two sets A and B as d(A, B) = size(A − B), where size is a function returning the number of elements in a set. This distance measure, which is an integer value greater than or equal to 0, does not satisfy the second part of the positivity property, the symmetry property, or the triangle inequality. However, these properties can be made to hold if the dissimilarity measure is modiﬁed as follows: d(A, B) = size(A − B) + size(B − A). See Exercise 21 on page 94. Example 2.15 (Nonmetric Dissimilarities: Time). This example gives a more everyday example of a dissimilarity measure that is not a metric, but that is still useful. Deﬁne a measure of the distance between times of the day as follows: if t1 ≤ t2 t2 − t1 d(t1 , t2 ) = . (2.4) 24 + (t2 − t1 ) if t1 ≥ t2 To illustrate, d (1PM, 2PM) = 1 hour, while d (2PM, 1PM) = 23 hours. Such a deﬁnition would make sense, for example, when answering the question: “If an event occurs at 1PM every day, and it is now 2PM, how long do I have to wait for that event to occur again?”
2.4.4
Similarities between Data Objects
For similarities, the triangle inequality (or the analogous property) typically does not hold, but symmetry and positivity typically do. To be explicit, if s(x, y) is the similarity between points x and y, then the typical properties of similarities are the following: 1. s(x, y) = 1 only if x = y. (0 ≤ s ≤ 1) 2. s(x, y) = s(y, x) for all x and y. (Symmetry) There is no general analog of the triangle inequality for similarity measures. It is sometimes possible, however, to show that a similarity measure can easily be converted to a metric distance. The cosine and Jaccard similarity measures, which are discussed shortly, are two examples. Also, for speciﬁc similarity measures, it is possible to derive mathematical bounds on the similarity between two objects that are similar in spirit to the triangle inequality.
72
2.4
Measures of Similarity and Dissimilarity
Example 2.16 (A Nonsymmetric Similarity Measure). Consider an experiment in which people are asked to classify a small set of characters as they ﬂash on a screen. The confusion matrix for this experiment records how often each character is classiﬁed as itself, and how often each is classiﬁed as another character. For instance, suppose that “0” appeared 200 times and was classiﬁed as a “0” 160 times, but as an “o” 40 times. Likewise, suppose that ‘o’ appeared 200 times and was classiﬁed as an “o” 170 times, but as “0” only 30 times. If we take these counts as a measure of the similarity between two characters, then we have a similarity measure, but one that is not symmetric. In such situations, the similarity measure is often made symmetric by setting s (x, y) = s (y, x) = (s(x, y)+s(y, x))/2, where s indicates the new similarity measure.
2.4.5
Examples of Proximity Measures
This section provides speciﬁc examples of some similarity and dissimilarity measures. Similarity Measures for Binary Data Similarity measures between objects that contain only binary attributes are called similarity coeﬃcients, and typically have values between 0 and 1. A value of 1 indicates that the two objects are completely similar, while a value of 0 indicates that the objects are not at all similar. There are many rationales for why one coeﬃcient is better than another in speciﬁc instances. Let x and y be two objects that consist of n binary attributes. The comparison of two such objects, i.e., two binary vectors, leads to the following four quantities (frequencies): f00 f01 f10 f11
= = = =
the the the the
number number number number
of of of of
attributes attributes attributes attributes
where where where where
x x x x
is is is is
0 0 1 1
and and and and
y y y y
is is is is
0 1 0 1
Simple Matching Coeﬃcient One commonly used similarity coeﬃcient is the simple matching coeﬃcient (SM C), which is deﬁned as SM C =
f11 + f00 number of matching attribute values = . number of attributes f01 + f10 + f11 + f00
(2.5)
73
Chapter 2
Data
This measure counts both presences and absences equally. Consequently, the SM C could be used to ﬁnd students who had answered questions similarly on a test that consisted only of true/false questions. Jaccard Coeﬃcient Suppose that x and y are data objects that represent two rows (two transactions) of a transaction matrix (see Section 2.1.2). If each asymmetric binary attribute corresponds to an item in a store, then a 1 indicates that the item was purchased, while a 0 indicates that the product was not purchased. Since the number of products not purchased by any customer far outnumbers the number of products that were purchased, a similarity measure such as SM C would say that all transactions are very similar. As a result, the Jaccard coeﬃcient is frequently used to handle objects consisting of asymmetric binary attributes. The Jaccard coeﬃcient, which is often symbolized by J, is given by the following equation: J=
f11 number of matching presences = . number of attributes not involved in 00 matches f01 + f10 + f11
(2.6)
Example 2.17 (The SMC and Jaccard Similarity Coeﬃcients). To illustrate the diﬀerence between these two similarity measures, we calculate SM C and J for the following two binary vectors. x = (1, 0, 0, 0, 0, 0, 0, 0, 0, 0) y = (0, 0, 0, 0, 0, 0, 1, 0, 0, 1) f01 f10 f00 f11
=2 =1 =7 =0
SM C = J=
the the the the
number number number number
of of of of
f11 +f00 f01 +f10 +f11 +f00
f11 f01 +f10 +f11
=
0 2+1+0
attributes attributes attributes attributes
=
0+7 2+1+0+7
where where where where
x x x x
was was was was
0 1 0 1
and and and and
y y y y
was was was was
1 0 0 1
= 0.7
=0
Cosine Similarity Documents are often represented as vectors, where each attribute represents the frequency with which a particular term (word) occurs in the document. It is more complicated than this, of course, since certain common words are ig
74
2.4
Measures of Similarity and Dissimilarity
nored and various processing techniques are used to account for diﬀerent forms of the same word, diﬀering document lengths, and diﬀerent word frequencies. Even though documents have thousands or tens of thousands of attributes (terms), each document is sparse since it has relatively few nonzero attributes. (The normalizations used for documents do not create a nonzero entry where there was a zero entry; i.e., they preserve sparsity.) Thus, as with transaction data, similarity should not depend on the number of shared 0 values since any two documents are likely to “not contain” many of the same words, and therefore, if 0–0 matches are counted, most documents will be highly similar to most other documents. Therefore, a similarity measure for documents needs to ignores 0–0 matches like the Jaccard measure, but also must be able to handle nonbinary vectors. The cosine similarity, deﬁned next, is one of the most common measure of document similarity. If x and y are two document vectors, then cos(x, y) =
x·y , x y
where · indicates the vector dot product, x · y =
√ n 2 length of vector x, x = x · x. k=1 xk =
(2.7)
n
k=1 xk yk ,
and x is the
Example 2.18 (Cosine Similarity of Two Document Vectors). This example calculates the cosine similarity for the following two data objects, which might represent document vectors: x = (3, 2, 0, 5, 0, 0, 0, 2, 0, 0) y = (1, 0, 0, 0, 0, 0, 0, 1, 0, 2) x·y =3∗1+2∗0+0∗0+5∗0+0∗0+0∗0+0∗0+2∗1+0∗0+0∗2=5 √ x = 3 ∗ 3 + 2 ∗ 2 + 0 ∗ 0 + 5 ∗ 5 + 0 ∗ 0 + 0 ∗ 0 + 0 ∗ 0 + 2 ∗ 2 + 0 ∗ 0 + 0 ∗ 0 = 6.48 √ y = 1 ∗ 1 + 0 ∗ 0 + 0 ∗ 0 + 0 ∗ 0 + 0 ∗ 0 + 0 ∗ 0 + 0 ∗ 0 + 1 ∗ 1 + 0 ∗ 0 + 2 ∗ 2 = 2.24 cos(x, y) = 0.31
As indicated by Figure 2.16, cosine similarity really is a measure of the (cosine of the) angle between x and y. Thus, if the cosine similarity is 1, the angle between x and y is 0◦ , and x and y are the same except for magnitude (length). If the cosine similarity is 0, then the angle between x and y is 90◦ , and they do not share any terms (words).
75
Chapter 2
Data
x
θ
y
Figure 2.16. Geometric illustration of the cosine measure.
Equation 2.7 can be written as Equation 2.8. cos(x, y) =
y x · = x · y , x y
(2.8)
where x = x/x and y = y/y. Dividing x and y by their lengths normalizes them to have a length of 1. This means that cosine similarity does not take the magnitude of the two data objects into account when computing similarity. (Euclidean distance might be a better choice when magnitude is important.) For vectors with a length of 1, the cosine measure can be calculated by taking a simple dot product. Consequently, when many cosine similarities between objects are being computed, normalizing the objects to have unit length can reduce the time required. Extended Jaccard Coeﬃcient (Tanimoto Coeﬃcient) The extended Jaccard coeﬃcient can be used for document data and that reduces to the Jaccard coeﬃcient in the case of binary attributes. The extended Jaccard coeﬃcient is also known as the Tanimoto coeﬃcient. (However, there is another coeﬃcient that is also known as the Tanimoto coeﬃcient.) This coeﬃcient, which we shall represent as EJ, is deﬁned by the following equation: EJ(x, y) =
x·y . x2 + y2 − x · y
(2.9)
Correlation The correlation between two data objects that have binary or continuous variables is a measure of the linear relationship between the attributes of the objects. (The calculation of correlation between attributes, which is more common, can be deﬁned similarly.) More precisely, Pearson’s correlation
76
2.4
Measures of Similarity and Dissimilarity
coeﬃcient between two data objects, x and y, is deﬁned by the following equation:
corr(x, y) =
sxy covariance(x, y) , (2.10) = standard deviation(x) ∗ standard deviation(y) sx sy
where we are using the following standard statistical notation and deﬁnitions: 1 (xk − x)(yk − y) n−1 n
covariance(x, y) = sxy =
(2.11)
k=1
standard deviation(x)
= sx =
standard deviation(y)
= sy =
1 (xk − x)2 n−1 n
k=1
1 (yk − y)2 n−1 n
k=1
1 xk is the mean of x n n
x = y
=
1 n
k=1 n
yk is the mean of y
k=1
Example 2.19 (Perfect Correlation). Correlation is always in the range −1 to 1. A correlation of 1 (−1) means that x and y have a perfect positive (negative) linear relationship; that is, xk = ayk + b, where a and b are constants. The following two sets of values for x and y indicate cases where the correlation is −1 and +1, respectively. In the ﬁrst case, the means of x and y were chosen to be 0, for simplicity. x = (−3, 6, 0, 3, −6) y = ( 1, −2, 0, −1, 2) x = (3, 6, 0, 3, 6) y = (1, 2, 0, 1, 2)
77
Chapter 2
Data
–1.00
–0.90
–0.80
–0.70
–0.60
0.50
–0.40
–0.30
–0.20
–0.10
0.00
0.10
0.20
0.30
0.40
0.50
0.60
0.70
0.80
0.90
1.00
Figure 2.17. Scatter plots illustrating correlations from −1 to 1.
Example 2.20 (Nonlinear Relationships). If the correlation is 0, then there is no linear relationship between the attributes of the two data objects. However, nonlinear relationships may still exist. In the following example, xk = yk2 , but their correlation is 0. x = (−3, −2, −1, 0, 1, 2, 3) y = ( 9, 4, 1, 0, 1, 4, 9) Example 2.21 (Visualizing Correlation). It is also easy to judge the correlation between two data objects x and y by plotting pairs of corresponding attribute values. Figure 2.17 shows a number of these plots when x and y have 30 attributes and the values of these attributes are randomly generated (with a normal distribution) so that the correlation of x and y ranges from −1 to 1. Each circle in a plot represents one of the 30 attributes; its x coordinate is the value of one of the attributes for x, while its y coordinate is the value of the same attribute for y. If we transform x and y by subtracting oﬀ their means and then normalizing them so that their lengths are 1, then their correlation can be calculated by
78
2.4
Measures of Similarity and Dissimilarity
taking the dot product. Notice that this is not the same as the standardization used in other contexts, where we make the transformations, xk = (xk − x)/sx and yk = (yk − y)/sy . Bregman Divergence∗ This section provides a brief description of Bregman divergences, which are a family of proximity functions that share some common properties. As a result, it is possible to construct general data mining algorithms, such as clustering algorithms, that work with any Bregman divergence. A concrete example is the Kmeans clustering algorithm (Section 8.2). Note that this section requires knowledge of vector calculus. Bregman divergences are loss or distortion functions. To understand the idea of a loss function, consider the following. Let x and y be two points, where y is regarded as the original point and x is some distortion or approximation of it. For example, x may be a point that was generated, for example, by adding random noise to y. The goal is to measure the resulting distortion or loss that results if y is approximated by x. Of course, the more similar x and y are, the smaller the loss or distortion. Thus, Bregman divergences can be used as dissimilarity functions. More formally, we have the following deﬁnition. Deﬁnition 2.6 (Bregman Divergence). Given a strictly convex function φ (with a few modest restrictions that are generally satisﬁed), the Bregman divergence (loss function) D(x, y) generated by that function is given by the following equation: D(x, y) = φ(x) − φ(y) − ∇φ(y), (x − y)
(2.12)
where ∇φ(y) is the gradient of φ evaluated at y, x − y, is the vector diﬀerence between x and y, and ∇φ(y), (x − y) is the inner product between ∇φ(x) and (x − y). For points in Euclidean space, the inner product is just the dot product. D(x, y) can be written as D(x, y) = φ(x) − L(x), where L(x) = φ(y) + ∇φ(y), (x − y) and represents the equation of a plane that is tangent to the function φ at y. Using calculus terminology, L(x) is the linearization of φ around the point y and the Bregman divergence is just the diﬀerence between a function and a linear approximation to that function. Diﬀerent Bregman divergences are obtained by using diﬀerent choices for φ. Example 2.22. We provide a concrete example using squared Euclidean distance, but restrict ourselves to one dimension to simplify the mathematics. Let
79
Chapter 2
Data
x and y be real numbers and φ(t) be the real valued function, φ(t) = t2 . In that case, the gradient reduces to the derivative and the dot product reduces to multiplication. Speciﬁcally, Equation 2.12 becomes Equation 2.13. D(x, y) = x2 − y 2 − 2y(x − y) = (x − y)2
(2.13)
The graph for this example, with y = 1, is shown in Figure 2.18. The Bregman divergence is shown for two values of x: x = 2 and x = 3. 10 9 8
D(3, 1)
7
y
6 5 φ(x) = x2 4
y = 2x –1
D(2, 1)
3 2 1 0 –4
–3
–2
–1
0 x
1
2
3
4
Figure 2.18. Illustration of Bregman divergence.
2.4.6
Issues in Proximity Calculation
This section discusses several important issues related to proximity measures: (1) how to handle the case in which attributes have diﬀerent scales and/or are correlated, (2) how to calculate proximity between objects that are composed of diﬀerent types of attributes, e.g., quantitative and qualitative, (3) and how to handle proximity calculation when attributes have diﬀerent weights; i.e., when not all attributes contribute equally to the proximity of objects.
80
2.4
Measures of Similarity and Dissimilarity
Standardization and Correlation for Distance Measures An important issue with distance measures is how to handle the situation when attributes do not have the same range of values. (This situation is often described by saying that “the variables have diﬀerent scales.”) Earlier, Euclidean distance was used to measure the distance between people based on two attributes: age and income. Unless these two attributes are standardized, the distance between two people will be dominated by income. A related issue is how to compute distance when there is correlation between some of the attributes, perhaps in addition to diﬀerences in the ranges of values. A generalization of Euclidean distance, the Mahalanobis distance, is useful when attributes are correlated, have diﬀerent ranges of values (different variances), and the distribution of the data is approximately Gaussian (normal). Speciﬁcally, the Mahalanobis distance between two objects (vectors) x and y is deﬁned as mahalanobis(x, y) = (x − y)Σ−1 (x − y)T ,
(2.14)
where Σ−1 is the inverse of the covariance matrix of the data. Note that the covariance matrix Σ is the matrix whose ij th entry is the covariance of the ith and j th attributes as deﬁned by Equation 2.11. Example 2.23. In Figure 2.19, there are 1000 points, whose x and y attributes have a correlation of 0.6. The distance between the two large points at the opposite ends of the long axis of the ellipse is 14.7 in terms of Euclidean distance, but only 6 with respect to Mahalanobis distance. In practice, computing the Mahalanobis distance is expensive, but can be worthwhile for data whose attributes are correlated. If the attributes are relatively uncorrelated, but have diﬀerent ranges, then standardizing the variables is suﬃcient.
Combining Similarities for Heterogeneous Attributes The previous deﬁnitions of similarity were based on approaches that assumed all the attributes were of the same type. A general approach is needed when the attributes are of diﬀerent types. One straightforward approach is to compute the similarity between each attribute separately using Table 2.7, and then combine these similarities using a method that results in a similarity between 0 and 1. Typically, the overall similarity is deﬁned as the average of all the individual attribute similarities.
81
Chapter 2
Data
5 4 3 2
y
1 0
–1 –2 –3 –4 –5 –8
–6
–4
–2
0 x
2
4
6
8
Figure 2.19. Set of twodimensional points. The Mahalanobis distance between the two points represented by large dots is 6; their Euclidean distance is 14.7.
Unfortunately, this approach does not work well if some of the attributes are asymmetric attributes. For example, if all the attributes are asymmetric binary attributes, then the similarity measure suggested previously reduces to the simple matching coeﬃcient, a measure that is not appropriate for asymmetric binary attributes. The easiest way to ﬁx this problem is to omit asymmetric attributes from the similarity calculation when their values are 0 for both of the objects whose similarity is being computed. A similar approach also works well for handling missing values. In summary, Algorithm 2.1 is eﬀective for computing an overall similarity between two objects, x and y, with diﬀerent types of attributes. This procedure can be easily modiﬁed to work with dissimilarities. Using Weights In much of the previous discussion, all attributes were treated equally when computing proximity. This is not desirable when some attributes are more important to the deﬁnition of proximity than others. To address these situations,
82
2.4
Measures of Similarity and Dissimilarity
Algorithm 2.1 Similarities of heterogeneous objects. 1: For the k th attribute, compute a similarity, sk (x, y), in the range [0, 1]. th 2: Deﬁne ⎧an indicator variable, δk , for the k attribute as follows:
0 if the k th attribute is an asymmetric attribute and ⎪ ⎪ ⎨ both objects have a value of 0, or if one of the objects δk = has a missing value for the k th attribute ⎪ ⎪ ⎩ 1 otherwise 3: Compute the overall similarity between the two objects using the following formula: n δ s (x, y) nk k similarity(x, y) = k=1 (2.15) k=1 δk
the formulas for proximity can be modiﬁed by weighting the contribution of each attribute. If the weights wk sum to 1, then (2.15) becomes n wk δk sk (x, y) similarity(x, y) = k=1n . (2.16) k=1 δk The deﬁnition of the Minkowski distance can also be modiﬁed as follows: d(x, y) =
n
1/r wk xk − yk 
r
.
(2.17)
k=1
2.4.7
Selecting the Right Proximity Measure
The following are a few general observations that may be helpful. First, the type of proximity measure should ﬁt the type of data. For many types of dense, continuous data, metric distance measures such as Euclidean distance are often used. Proximity between continuous attributes is most often expressed in terms of diﬀerences, and distance measures provide a welldeﬁned way of combining these diﬀerences into an overall proximity measure. Although attributes can have diﬀerent scales and be of diﬀering importance, these issues can often be dealt with as described earlier. For sparse data, which often consists of asymmetric attributes, we typically employ similarity measures that ignore 0–0 matches. Conceptually, this reﬂects the fact that, for a pair of complex objects, similarity depends on the number of characteristics they both share, rather than the number of characteristics they both lack. More speciﬁcally, for sparse, asymmetric data, most
83
Chapter 2
Data
objects have only a few of the characteristics described by the attributes, and thus, are highly similar in terms of the characteristics they do not have. The cosine, Jaccard, and extended Jaccard measures are appropriate for such data. There are other characteristics of data vectors that may need to be considered. Suppose, for example, that we are interested in comparing time series. If the magnitude of the time series is important (for example, each time series represent total sales of the same organization for a diﬀerent year), then we could use Euclidean distance. If the time series represent diﬀerent quantities (for example, blood pressure and oxygen consumption), then we usually want to determine if the time series have the same shape, not the same magnitude. Correlation, which uses a builtin normalization that accounts for diﬀerences in magnitude and level, would be more appropriate. In some cases, transformation or normalization of the data is important for obtaining a proper similarity measure since such transformations are not always present in proximity measures. For instance, time series may have trends or periodic patterns that signiﬁcantly impact similarity. Also, a proper computation of similarity may require that time lags be taken into account. Finally, two time series may only be similar over speciﬁc periods of time. For example, there is a strong relationship between temperature and the use of natural gas, but only during the heating season. Practical consideration can also be important. Sometimes, a one or more proximity measures are already in use in a particular ﬁeld, and thus, others will have answered the question of which proximity measures should be used. Other times, the software package or clustering algorithm being used may drastically limit the choices. If eﬃciency is a concern, then we may want to choose a proximity measure that has a property, such as the triangle inequality, that can be used to reduce the number of proximity calculations. (See Exercise 25.) However, if common practice or practical restrictions do not dictate a choice, then the proper choice of a proximity measure can be a timeconsuming task that requires careful consideration of both domain knowledge and the purpose for which the measure is being used. A number of diﬀerent similarity measures may need to be evaluated to see which ones produce results that make the most sense.
2.5
Bibliographic Notes
It is essential to understand the nature of the data that is being analyzed, and at a fundamental level, this is the subject of measurement theory. In
84
2.5
Bibliographic Notes
particular, one of the initial motivations for deﬁning types of attributes was to be precise about which statistical operations were valid for what sorts of data. We have presented the view of measurement theory that was initially described in a classic paper by S. S. Stevens [79]. (Tables 2.2 and 2.3 are derived from those presented by Stevens [80].) While this is the most common view and is reasonably easy to understand and apply, there is, of course, much more to measurement theory. An authoritative discussion can be found in a threevolume series on the foundations of measurement theory [63, 69, 81]. Also of interest is a wideranging article by Hand [55], which discusses measurement theory and statistics, and is accompanied by comments from other researchers in the ﬁeld. Finally, there are many books and articles that describe measurement issues for particular areas of science and engineering. Data quality is a broad subject that spans every discipline that uses data. Discussions of precision, bias, accuracy, and signiﬁcant ﬁgures can be found in many introductory science, engineering, and statistics textbooks. The view of data quality as “ﬁtness for use” is explained in more detail in the book by Redman [76]. Those interested in data quality may also be interested in MIT’s Total Data Quality Management program [70, 84]. However, the knowledge needed to deal with speciﬁc data quality issues in a particular domain is often best obtained by investigating the data quality practices of researchers in that ﬁeld. Aggregation is a less welldeﬁned subject than many other preprocessing tasks. However, aggregation is one of the main techniques used by the database area of Online Analytical Processing (OLAP), which is discussed in Chapter 3. There has also been relevant work in the area of symbolic data analysis (Bock and Diday [47]). One of the goals in this area is to summarize traditional record data in terms of symbolic data objects whose attributes are more complex than traditional attributes. Speciﬁcally, these attributes can have values that are sets of values (categories), intervals, or sets of values with weights (histograms). Another goal of symbolic data analysis is to be able to perform clustering, classiﬁcation, and other kinds of data analysis on data that consists of symbolic data objects. Sampling is a subject that has been well studied in statistics and related ﬁelds. Many introductory statistics books, such as the one by Lindgren [65], have some discussion on sampling, and there are entire books devoted to the subject, such as the classic text by Cochran [49]. A survey of sampling for data mining is provided by Gu and Liu [54], while a survey of sampling for databases is provided by Olken and Rotem [72]. There are a number of other data mining and databaserelated sampling references that may be of interest,
85
Chapter 2
Data
including papers by Palmer and Faloutsos [74], Provost et al. [75], Toivonen [82], and Zaki et al. [85]. In statistics, the traditional techniques that have been used for dimensionality reduction are multidimensional scaling (MDS) (Borg and Groenen [48], Kruskal and Uslaner [64]) and principal component analysis (PCA) (Jolliﬀe [58]), which is similar to singular value decomposition (SVD) (Demmel [50]). Dimensionality reduction is discussed in more detail in Appendix B. Discretization is a topic that has been extensively investigated in data mining. Some classiﬁcation algorithms only work with categorical data, and association analysis requires binary data, and thus, there is a signiﬁcant motivation to investigate how to best binarize or discretize continuous attributes. For association analysis, we refer the reader to work by Srikant and Agrawal [78], while some useful references for discretization in the area of classiﬁcation include work by Dougherty et al. [51], Elomaa and Rousu [52], Fayyad and Irani [53], and Hussain et al. [56]. Feature selection is another topic well investigated in data mining. A broad coverage of this topic is provided in a survey by Molina et al. [71] and two books by Liu and Motada [66, 67]. Other useful papers include those by Blum and Langley [46], Kohavi and John [62], and Liu et al. [68]. It is diﬃcult to provide references for the subject of feature transformations because practices vary from one discipline to another. Many statistics books have a discussion of transformations, but typically the discussion is restricted to a particular purpose, such as ensuring the normality of a variable or making sure that variables have equal variance. We oﬀer two references: Osborne [73] and Tukey [83]. While we have covered some of the most commonly used distance and similarity measures, there are hundreds of such measures and more are being created all the time. As with so many other topics in this chapter, many of these measures are speciﬁc to particular ﬁelds; e.g., in the area of time series see papers by Kalpakis et al. [59] and Keogh and Pazzani [61]. Clustering books provide the best general discussions. In particular, see the books by Anderberg [45], Jain and Dubes [57], Kaufman and Rousseeuw [60], and Sneath and Sokal [77].
Bibliography [45] M. R. Anderberg. Cluster Analysis for Applications. Academic Press, New York, December 1973. [46] A. Blum and P. Langley. Selection of Relevant Features and Examples in Machine Learning. Artificial Intelligence, 97(1–2):245–271, 1997.
86
Bibliography [47] H. H. Bock and E. Diday. Analysis of Symbolic Data: Exploratory Methods for Extracting Statistical Information from Complex Data (Studies in Classification, Data Analysis, and Knowledge Organization). SpringerVerlag Telos, January 2000. [48] I. Borg and P. Groenen. Modern Multidimensional Scaling—Theory and Applications. SpringerVerlag, February 1997. [49] W. G. Cochran. Sampling Techniques. John Wiley & Sons, 3rd edition, July 1977. [50] J. W. Demmel. Applied Numerical Linear Algebra. Society for Industrial & Applied Mathematics, September 1997. [51] J. Dougherty, R. Kohavi, and M. Sahami. Supervised and Unsupervised Discretization of Continuous Features. In Proc. of the 12th Intl. Conf. on Machine Learning, pages 194–202, 1995. [52] T. Elomaa and J. Rousu. General and Eﬃcient Multisplitting of Numerical Attributes. Machine Learning, 36(3):201–244, 1999. [53] U. M. Fayyad and K. B. Irani. Multiinterval discretization of continuousvalued attributes for classiﬁcation learning. In Proc. 13th Int. Joint Conf. on Artificial Intelligence, pages 1022–1027. Morgan Kaufman, 1993. [54] F. H. Gaohua Gu and H. Liu. Sampling and Its Application in Data Mining: A Survey. Technical Report TRA6/00, National University of Singapore, Singapore, 2000. [55] D. J. Hand. Statistics and the Theory of Measurement. Journal of the Royal Statistical Society: Series A (Statistics in Society), 159(3):445–492, 1996. [56] F. Hussain, H. Liu, C. L. Tan, and M. Dash. TRC6/99: Discretization: an enabling technique. Technical report, National University of Singapore, Singapore, 1999. [57] A. K. Jain and R. C. Dubes. Algorithms for Clustering Data. Prentice Hall Advanced Reference Series. Prentice Hall, March 1988. Book available online at http://www.cse.msu.edu/∼jain/Clustering Jain Dubes.pdf. [58] I. T. Jolliﬀe. Principal Component Analysis. Springer Verlag, 2nd edition, October 2002. [59] K. Kalpakis, D. Gada, and V. Puttagunta. Distance Measures for Eﬀective Clustering of ARIMA TimeSeries. In Proc. of the 2001 IEEE Intl. Conf. on Data Mining, pages 273–280. IEEE Computer Society, 2001. [60] L. Kaufman and P. J. Rousseeuw. Finding Groups in Data: An Introduction to Cluster Analysis. Wiley Series in Probability and Statistics. John Wiley and Sons, New York, November 1990. [61] E. J. Keogh and M. J. Pazzani. Scaling up dynamic time warping for datamining applications. In KDD, pages 285–289, 2000. [62] R. Kohavi and G. H. John. Wrappers for Feature Subset Selection. Artificial Intelligence, 97(1–2):273–324, 1997. [63] D. Krantz, R. D. Luce, P. Suppes, and A. Tversky. Foundations of Measurements: Volume 1: Additive and polynomial representations. Academic Press, New York, 1971. [64] J. B. Kruskal and E. M. Uslaner. Multidimensional Scaling. Sage Publications, August 1978. [65] B. W. Lindgren. Statistical Theory. CRC Press, January 1993. [66] H. Liu and H. Motoda, editors. Feature Extraction, Construction and Selection: A Data Mining Perspective. Kluwer International Series in Engineering and Computer Science, 453. Kluwer Academic Publishers, July 1998. [67] H. Liu and H. Motoda. Feature Selection for Knowledge Discovery and Data Mining. Kluwer International Series in Engineering and Computer Science, 454. Kluwer Academic Publishers, July 1998.
87
Chapter 2
Data
[68] H. Liu, H. Motoda, and L. Yu. Feature Extraction, Selection, and Construction. In N. Ye, editor, The Handbook of Data Mining, pages 22–41. Lawrence Erlbaum Associates, Inc., Mahwah, NJ, 2003. [69] R. D. Luce, D. Krantz, P. Suppes, and A. Tversky. Foundations of Measurements: Volume 3: Representation, Axiomatization, and Invariance. Academic Press, New York, 1990. [70] MIT Total Data Quality Management Program. web.mit.edu/tdqm/www/index.shtml, 2003. [71] L. C. Molina, L. Belanche, and A. Nebot. Feature Selection Algorithms: A Survey and Experimental Evaluation. In Proc. of the 2002 IEEE Intl. Conf. on Data Mining, 2002. [72] F. Olken and D. Rotem. Random Sampling from Databases—A Survey. Statistics & Computing, 5(1):25–42, March 1995. [73] J. Osborne. Notes on the Use of Data Transformations. Practical Assessment, Research & Evaluation, 28(6), 2002. [74] C. R. Palmer and C. Faloutsos. Density biased sampling: An improved method for data mining and clustering. ACM SIGMOD Record, 29(2):82–92, 2000. [75] F. J. Provost, D. Jensen, and T. Oates. Eﬃcient Progressive Sampling. In Proc. of the 5th Intl. Conf. on Knowledge Discovery and Data Mining, pages 23–32, 1999. [76] T. C. Redman. Data Quality: The Field Guide. Digital Press, January 2001. [77] P. H. A. Sneath and R. R. Sokal. Numerical Taxonomy. Freeman, San Francisco, 1971. [78] R. Srikant and R. Agrawal. Mining Quantitative Association Rules in Large Relational Tables. In Proc. of 1996 ACMSIGMOD Intl. Conf. on Management of Data, pages 1–12, Montreal, Quebec, Canada, August 1996. [79] S. S. Stevens. On the Theory of Scales of Measurement. Science, 103(2684):677–680, June 1946. [80] S. S. Stevens. Measurement. In G. M. Maranell, editor, Scaling: A Sourcebook for Behavioral Scientists, pages 22–41. Aldine Publishing Co., Chicago, 1974. [81] P. Suppes, D. Krantz, R. D. Luce, and A. Tversky. Foundations of Measurements: Volume 2: Geometrical, Threshold, and Probabilistic Representations. Academic Press, New York, 1989. [82] H. Toivonen. Sampling Large Databases for Association Rules. In VLDB96, pages 134–145. Morgan Kaufman, September 1996. [83] J. W. Tukey. On the Comparative Anatomy of Transformations. Annals of Mathematical Statistics, 28(3):602–632, September 1957. [84] R. Y. Wang, M. Ziad, Y. W. Lee, and Y. R. Wang. Data Quality. The Kluwer International Series on Advances in Database Systems, Volume 23. Kluwer Academic Publishers, January 2001. [85] M. J. Zaki, S. Parthasarathy, W. Li, and M. Ogihara. Evaluation of Sampling for Data Mining of Association Rules. Technical Report TR617, Rensselaer Polytechnic Institute, 1996.
2.6
Exercises
1. In the initial example of Chapter 2, the statistician says, “Yes, ﬁelds 2 and 3 are basically the same.” Can you tell from the three lines of sample data that are shown why she says that?
88
2.6
Exercises
2. Classify the following attributes as binary, discrete, or continuous. Also classify them as qualitative (nominal or ordinal) or quantitative (interval or ratio). Some cases may have more than one interpretation, so brieﬂy indicate your reasoning if you think there may be some ambiguity. Example: Age in years. Answer: Discrete, quantitative, ratio (a) Time in terms of AM or PM. (b) Brightness as measured by a light meter. (c) Brightness as measured by people’s judgments. (d) Angles as measured in degrees between 0 and 360. (e) Bronze, Silver, and Gold medals as awarded at the Olympics. (f) Height above sea level. (g) Number of patients in a hospital. (h) ISBN numbers for books. (Look up the format on the Web.) (i) Ability to pass light in terms of the following values: opaque, translucent, transparent. (j) Military rank. (k) Distance from the center of campus. (l) Density of a substance in grams per cubic centimeter. (m) Coat check number. (When you attend an event, you can often give your coat to someone who, in turn, gives you a number that you can use to claim your coat when you leave.) 3. You are approached by the marketing director of a local company, who believes that he has devised a foolproof way to measure customer satisfaction. He explains his scheme as follows: “It’s so simple that I can’t believe that no one has thought of it before. I just keep track of the number of customer complaints for each product. I read in a data mining book that counts are ratio attributes, and so, my measure of product satisfaction must be a ratio attribute. But when I rated the products based on my new customer satisfaction measure and showed them to my boss, he told me that I had overlooked the obvious, and that my measure was worthless. I think that he was just mad because our bestselling product had the worst satisfaction since it had the most complaints. Could you help me set him straight?” (a) Who is right, the marketing director or his boss? If you answered, his boss, what would you do to ﬁx the measure of satisfaction? (b) What can you say about the attribute type of the original product satisfaction attribute?
89
Chapter 2
Data
4. A few months later, you are again approached by the same marketing director as in Exercise 3. This time, he has devised a better approach to measure the extent to which a customer prefers one product over other, similar products. He explains, “When we develop new products, we typically create several variations and evaluate which one customers prefer. Our standard procedure is to give our test subjects all of the product variations at one time and then ask them to rank the product variations in order of preference. However, our test subjects are very indecisive, especially when there are more than two products. As a result, testing takes forever. I suggested that we perform the comparisons in pairs and then use these comparisons to get the rankings. Thus, if we have three product variations, we have the customers compare variations 1 and 2, then 2 and 3, and ﬁnally 3 and 1. Our testing time with my new procedure is a third of what it was for the old procedure, but the employees conducting the tests complain that they cannot come up with a consistent ranking from the results. And my boss wants the latest product evaluations, yesterday. I should also mention that he was the person who came up with the old product evaluation approach. Can you help me?” (a) Is the marketing director in trouble? Will his approach work for generating an ordinal ranking of the product variations in terms of customer preference? Explain. (b) Is there a way to ﬁx the marketing director’s approach? More generally, what can you say about trying to create an ordinal measurement scale based on pairwise comparisons? (c) For the original product evaluation scheme, the overall rankings of each product variation are found by computing its average over all test subjects. Comment on whether you think that this is a reasonable approach. What other approaches might you take? 5. Can you think of a situation in which identiﬁcation numbers would be useful for prediction? 6. An educational psychologist wants to use association analysis to analyze test results. The test consists of 100 questions with four possible answers each. (a) How would you convert this data into a form suitable for association analysis? (b) In particular, what type of attributes would you have and how many of them are there? 7. Which of the following quantities is likely to show more temporal autocorrelation: daily rainfall or daily temperature? Why? 8. Discuss why a documentterm matrix is an example of a data set that has asymmetric discrete or asymmetric continuous features.
90
2.6
Exercises
9. Many sciences rely on observation instead of (or in addition to) designed experiments. Compare the data quality issues involved in observational science with those of experimental science and data mining. 10. Discuss the diﬀerence between the precision of a measurement and the terms single and double precision, as they are used in computer science, typically to represent ﬂoatingpoint numbers that require 32 and 64 bits, respectively. 11. Give at least two advantages to working with data stored in text ﬁles instead of in a binary format. 12. Distinguish between noise and outliers. Be sure to consider the following questions. (a) Is noise ever interesting or desirable? Outliers? (b) Can noise objects be outliers? (c) Are noise objects always outliers? (d) Are outliers always noise objects? (e) Can noise make a typical value into an unusual one, or vice versa? 13. Consider the problem of ﬁnding the K nearest neighbors of a data object. A programmer designs Algorithm 2.2 for this task.
Algorithm 2.2 Algorithm for ﬁnding K nearest neighbors. 1: for i = 1 to number of data objects do 2: Find the distances of the ith object to all other objects. 3: Sort these distances in decreasing order.
(Keep track of which object is associated with each distance.) 4: return the objects associated with the ﬁrst K distances of the sorted list 5: end for
(a) Describe the potential problems with this algorithm if there are duplicate objects in the data set. Assume the distance function will only return a distance of 0 for objects that are the same. (b) How would you ﬁx this problem? 14. The following attributes are measured for members of a herd of Asian elephants: weight, height, tusk length, trunk length, and ear area. Based on these measurements, what sort of similarity measure from Section 2.4 would you use to compare or group these elephants? Justify your answer and explain any special circumstances.
91
Chapter 2
Data
15. You are given a set of m objects that is divided into K groups, where the ith group is of size mi . If the goal is to obtain a sample of size n < m, what is the diﬀerence between the following two sampling schemes? (Assume sampling with replacement.) (a) We randomly select n ∗ mi /m elements from each group. (b) We randomly select n elements from the data set, without regard for the group to which an object belongs. 16. Consider a documentterm matrix, where tfij is the frequency of the ith word (term) in the j th document and m is the number of documents. Consider the variable transformation that is deﬁned by m = tfij ∗ log , (2.18) tfij dfi where dfi is the number of documents in which the ith term appears, which is known as the document frequency of the term. This transformation is known as the inverse document frequency transformation. (a) What is the eﬀect of this transformation if a term occurs in one document? In every document? (b) What might be the purpose of this transformation? 17. Assume that we apply a square root transformation to a ratio attribute x to obtain the new attribute x∗ . As part of your analysis, you identify an interval (a, b) in which x∗ has a linear relationship to another attribute y. (a) What is the corresponding interval (a, b) in terms of x? (b) Give an equation that relates y to x. 18. This exercise compares and contrasts some similarity and distance measures. (a) For binary data, the L1 distance corresponds to the Hamming distance; that is, the number of bits that are diﬀerent between two binary vectors. The Jaccard similarity is a measure of the similarity between two binary vectors. Compute the Hamming distance and the Jaccard similarity between the following two binary vectors. x = 0101010001 y = 0100011000
(b) Which approach, Jaccard or Hamming distance, is more similar to the Simple Matching Coeﬃcient, and which approach is more similar to the cosine measure? Explain. (Note: The Hamming measure is a distance, while the other three measures are similarities, but don’t let this confuse you.)
92
2.6
Exercises
(c) Suppose that you are comparing how similar two organisms of diﬀerent species are in terms of the number of genes they share. Describe which measure, Hamming or Jaccard, you think would be more appropriate for comparing the genetic makeup of two organisms. Explain. (Assume that each animal is represented as a binary vector, where each attribute is 1 if a particular gene is present in the organism and 0 otherwise.) (d) If you wanted to compare the genetic makeup of two organisms of the same species, e.g., two human beings, would you use the Hamming distance, the Jaccard coeﬃcient, or a diﬀerent measure of similarity or distance? Explain. (Note that two human beings share > 99.9% of the same genes.) 19. For the following vectors, x and y, calculate the indicated similarity or distance measures. (a) x = (1, 1, 1, 1), y = (2, 2, 2, 2) cosine, correlation, Euclidean (b) x = (0, 1, 0, 1), y = (1, 0, 1, 0) cosine, correlation, Euclidean, Jaccard (c) x = (0, −1, 0, 1), y = (1, 0, −1, 0) cosine, correlation, Euclidean (d) x = (1, 1, 0, 1, 0, 1), y = (1, 1, 1, 0, 0, 1) cosine, correlation, Jaccard (e) x = (2, −1, 0, 2, 0, −3), y = (−1, 1, −1, 0, 0, −1) cosine, correlation 20. Here, we further explore the cosine and correlation measures. (a) What is the range of values that are possible for the cosine measure? (b) If two objects have a cosine measure of 1, are they identical? Explain. (c) What is the relationship of the cosine measure to correlation, if any? (Hint: Look at statistical measures such as mean and standard deviation in cases where cosine and correlation are the same and diﬀerent.) (d) Figure 2.20(a) shows the relationship of the cosine measure to Euclidean distance for 100,000 randomly generated points that have been normalized to have an L2 length of 1. What general observation can you make about the relationship between Euclidean distance and cosine similarity when vectors have an L2 norm of 1? (e) Figure 2.20(b) shows the relationship of correlation to Euclidean distance for 100,000 randomly generated points that have been standardized to have a mean of 0 and a standard deviation of 1. What general observation can you make about the relationship between Euclidean distance and correlation when the vectors have been standardized to have a mean of 0 and a standard deviation of 1? (f) Derive the mathematical relationship between cosine similarity and Euclidean distance when each data object has an L2 length of 1. (g) Derive the mathematical relationship between correlation and Euclidean distance when each data point has been been standardized by subtracting its mean and dividing by its standard deviation.
93
Data
1.4
1.4
1.2
1.2 Euclidean Distance
Euclidean Distance
Chapter 2
1 0.8 0.6 0.4 0.2 0 0
1 0.8 0.6 0.4 0.2
0.2
0.4 0.6 Cosine Similarity
0.8
(a) Relationship between Euclidean distance and the cosine measure.
1
0 0
0.2
0.4 0.6 Correlation
0.8
1
(b) Relationship between Euclidean distance and correlation.
Figure 2.20. Graphs for Exercise 20.
21. Show that the set diﬀerence metric given by d(A, B) = size(A − B) + size(B − A)
(2.19)
satisﬁes the metric axioms given on page 70. A and B are sets and A − B is the set diﬀerence. 22. Discuss how you might map correlation values from the interval [−1,1] to the interval [0,1]. Note that the type of transformation that you use might depend on the application that you have in mind. Thus, consider two applications: clustering time series and predicting the behavior of one time series given another. 23. Given a similarity measure with values in the interval [0,1] describe two ways to transform this similarity value into a dissimilarity value in the interval [0,∞]. 24. Proximity is typically deﬁned between a pair of objects. (a) Deﬁne two ways in which you might deﬁne the proximity among a group of objects. (b) How might you deﬁne the distance between two sets of points in Euclidean space? (c) How might you deﬁne the proximity between two sets of data objects? (Make no assumption about the data objects, except that a proximity measure is deﬁned between any pair of objects.) 25. You are given a set of points S in Euclidean space, as well as the distance of each point in S to a point x. (It does not matter if x ∈ S.)
94
2.6
Exercises
(a) If the goal is to ﬁnd all points within a speciﬁed distance ε of point y, y = x, explain how you could use the triangle inequality and the already calculated distances to x to potentially reduce the number of distance calculations necessary? Hint: The triangle inequality, d(x, z) ≤ d(x, y) + d(y, x), can be rewritten as d(x, y) ≥ d(x, z) − d(y, z). (b) In general, how would the distance between x and y aﬀect the number of distance calculations? (c) Suppose that you can ﬁnd a small subset of points S , from the original data set, such that every point in the data set is within a speciﬁed distance ε of at least one of the points in S , and that you also have the pairwise distance matrix for S . Describe a technique that uses this information to compute, with a minimum of distance calculations, the set of all points within a distance of β of a speciﬁed point from the data set. 26. Show that 1 minus the Jaccard similarity is a distance measure between two data objects, x and y, that satisﬁes the metric axioms given on page 70. Speciﬁcally, d(x, y) = 1 − J(x, y). 27. Show that the distance measure deﬁned as the angle between two data vectors, x and y, satisﬁes the metric axioms given on page 70. Speciﬁcally, d(x, y) = arccos(cos(x, y)). 28. Explain why computing the proximity between two attributes is often simpler than computing the similarity between two objects.
95
96
3 Exploring Data The previous chapter addressed highlevel data issues that are important in the knowledge discovery process. This chapter provides an introduction to data exploration, which is a preliminary investigation of the data in order to better understand its speciﬁc characteristics. Data exploration can aid in selecting the appropriate preprocessing and data analysis techniques. It can even address some of the questions typically answered by data mining. For example, patterns can sometimes be found by visually inspecting the data. Also, some of the techniques used in data exploration, such as visualization, can be used to understand and interpret data mining results. This chapter covers three major topics: summary statistics, visualization, and OnLine Analytical Processing (OLAP). Summary statistics, such as the mean and standard deviation of a set of values, and visualization techniques, such as histograms and scatter plots, are standard methods that are widely employed for data exploration. OLAP, which is a more recent development, consists of a set of techniques for exploring multidimensional arrays of values. OLAPrelated analysis functions focus on various ways to create summary data tables from a multidimensional data array. These techniques include aggregating data either across various dimensions or across various attribute values. For instance, if we are given sales information reported according to product, location, and date, OLAP techniques can be used to create a summary that describes the sales activity at a particular location by month and product category. The topics covered in this chapter have considerable overlap with the area known as Exploratory Data Analysis (EDA), which was created in the 1970s by the prominent statistician, John Tukey. This chapter, like EDA, places a heavy emphasis on visualization. Unlike EDA, this chapter does not include topics such as cluster analysis or anomaly detection. There are two
From Chapter 3 of Introduction to Data Mining, First Edition. PangNing Tan, Michael Steinbach, Vipin Kumar. Copyright © 2006 by Pearson Education, Inc. All rights reserved.
97
Chapter 3
Exploring Data
reasons for this. First, data mining views descriptive data analysis techniques as an end in themselves, whereas statistics, from which EDA originated, tends to view hypothesisbased testing as the ﬁnal goal. Second, cluster analysis and anomaly detection are large areas and require full chapters for an indepth discussion. Hence, cluster analysis is covered in Chapters 8 and 9, while anomaly detection is discussed in Chapter 10.
3.1
The Iris Data Set
In the following discussion, we will often refer to the Iris data set that is available from the University of California at Irvine (UCI) Machine Learning Repository. It consists of information on 150 Iris ﬂowers, 50 each from one of three Iris species: Setosa, Versicolour, and Virginica. Each ﬂower is characterized by ﬁve attributes: 1. sepal length in centimeters 2. sepal width in centimeters 3. petal length in centimeters 4. petal width in centimeters 5. class (Setosa, Versicolour, Virginica) The sepals of a ﬂower are the outer structures that protect the more fragile parts of the ﬂower, such as the petals. In many ﬂowers, the sepals are green, and only the petals are colorful. For Irises, however, the sepals are also colorful. As illustrated by the picture of a Virginica Iris in Figure 3.1, the sepals of an Iris are larger than the petals and are drooping, while the petals are upright.
3.2
Summary Statistics
Summary statistics are quantities, such as the mean and standard deviation, that capture various characteristics of a potentially large set of values with a single number or a small set of numbers. Everyday examples of summary statistics are the average household income or the fraction of college students who complete an undergraduate degree in four years. Indeed, for many people, summary statistics are the most visible manifestation of statistics. We will concentrate on summary statistics for the values of a single attribute, but will provide a brief description of some multivariate summary statistics.
98
3.2
Summary Statistics
Figure 3.1. Picture of Iris Virginica. Robert H. Mohlenbrock @ USDANRCS PLANTS Database/ USDA NRCS. 1995. Northeast wetland flora: Field office guide to plant species. Northeast National Technical Center, Chester, PA. Background removed.
This section considers only the descriptive nature of summary statistics. However, as described in Appendix C, statistics views data as arising from an underlying statistical process that is characterized by various parameters, and some of the summary statistics discussed here can be viewed as estimates of statistical parameters of the underlying distribution that generated the data.
3.2.1
Frequencies and the Mode
Given a set of unordered categorical values, there is not much that can be done to further characterize the values except to compute the frequency with which each value occurs for a particular set of data. Given a categorical attribute x, which can take values {v1 , . . . , vi , . . . vk } and a set of m objects, the frequency of a value vi is deﬁned as frequency(vi ) =
number of objects with attribute value vi . m
(3.1)
The mode of a categorical attribute is the value that has the highest frequency.
99
Chapter 3
Exploring Data
Example 3.1. Consider a set of students who have an attribute, class, which can take values from the set {f reshman, sophomore, junior, senior}. Table 3.1 shows the number of students for each value of the class attribute. The mode of the class attribute is f reshman, with a frequency of 0.33. This may indicate dropouts due to attrition or a larger than usual freshman class. Table 3.1. Class size for students in a hypothetical college. Class freshman sophomore junior senior
Size 140 160 130 170
Frequency 0.33 0.27 0.22 0.18
Categorical attributes often, but not always, have a small number of values, and consequently, the mode and frequencies of these values can be interesting and useful. Notice, though, that for the Iris data set and the class attribute, the three types of ﬂower all have the same frequency, and therefore, the notion of a mode is not interesting. For continuous data, the mode, as currently deﬁned, is often not useful because a single value may not occur more than once. Nonetheless, in some cases, the mode may indicate important information about the nature of the values or the presence of missing values. For example, the heights of 20 people measured to the nearest millimeter will typically not repeat, but if the heights are measured to the nearest tenth of a meter, then some people may have the same height. Also, if a unique value is used to indicate a missing value, then this value will often show up as the mode.
3.2.2
Percentiles
For ordered data, it is more useful to consider the percentiles of a set of values. In particular, given an ordinal or continuous attribute x and a number p between 0 and 100, the pth percentile xp is a value of x such that p% of the observed values of x are less than xp . For instance, the 50th percentile is the value x50% such that 50% of all values of x are less than x50% . Table 3.2 shows the percentiles for the four quantitative attributes of the Iris data set.
100
3.2
Summary Statistics
Table 3.2. Percentiles for sepal length, sepal width, petal length, and petal width. (All values are in centimeters.) Percentile 0 10 20 30 40 50 60 70 80 90 100
Sepal Length 4.3 4.8 5.0 5.2 5.6 5.8 6.1 6.3 6.6 6.9 7.9
Sepal Width 2.0 2.5 2.7 2.8 3.0 3.0 3.1 3.2 3.4 3.6 4.4
Petal Length 1.0 1.4 1.5 1.7 3.9 4.4 4.6 5.0 5.4 5.8 6.9
Petal Width 0.1 0.2 0.2 0.4 1.2 1.3 1.5 1.8 1.9 2.2 2.5
Example 3.2. The percentiles, x0% , x10% , . . . , x90% , x100% of the integers from 1 to 10 are, in order, the following: 1.0, 1.5, 2.5, 3.5, 4.5, 5.5, 6.5, 7.5, 8.5, 9.5, 10.0. By tradition, min(x) = x0% and max(x) = x100% .
3.2.3
Measures of Location: Mean and Median
For continuous data, two of the most widely used summary statistics are the mean and median, which are measures of the location of a set of values. Consider a set of m objects and an attribute x. Let {x1 , . . . , xm } be the attribute values of x for these m objects. As a concrete example, these values might be the heights of m children. Let {x(1) , . . . , x(m) } represent the values of x after they have been sorted in nondecreasing order. Thus, x(1) = min(x) and x(m) = max(x). Then, the mean and median are deﬁned as follows: 1 xi m m
mean(x) = x =
(3.2)
i=1
median(x) =
if m is odd, i.e., m = 2r + 1 x(r+1) 1 2 (x(r) + x(r+1) ) if m is even, i.e., m = 2r
(3.3)
To summarize, the median is the middle value if there are an odd number of values, and the average of the two middle values if the number of values is even. Thus, for seven values, the median is x(4) , while for ten values, the median is 12 (x(5) + x(6) ).
101
Chapter 3
Exploring Data
Although the mean is sometimes interpreted as the middle of a set of values, this is only correct if the values are distributed in a symmetric manner. If the distribution of values is skewed, then the median is a better indicator of the middle. Also, the mean is sensitive to the presence of outliers. For data with outliers, the median again provides a more robust estimate of the middle of a set of values. To overcome problems with the traditional deﬁnition of a mean, the notion of a trimmed mean is sometimes used. A percentage p between 0 and 100 is speciﬁed, the top and bottom (p/2)% of the data is thrown out, and the mean is then calculated in the normal way. The median is a trimmed mean with p = 100%, while the standard mean corresponds to p = 0%. Example 3.3. Consider the set of values {1, 2, 3, 4, 5, 90}. The mean of these values is 17.5, while the median is 3.5. The trimmed mean with p = 40% is also 3.5. Example 3.4. The means, medians, and trimmed means (p = 20%) of the four quantitative attributes of the Iris data are given in Table 3.3. The three measures of location have similar values except for the attribute petal length. Table 3.3. Means and medians for sepal length, sepal width, petal length, and petal width. (All values are in centimeters.) Measure mean median trimmed mean (20%)
3.2.4
Sepal Length 5.84 5.80 5.79
Sepal Width 3.05 3.00 3.02
Petal Length 3.76 4.35 3.72
Petal Width 1.20 1.30 1.12
Measures of Spread: Range and Variance
Another set of commonly used summary statistics for continuous data are those that measure the dispersion or spread of a set of values. Such measures indicate if the attribute values are widely spread out or if they are relatively concentrated around a single point such as the mean. The simplest measure of spread is the range, which, given an attribute x with a set of m values {x1 , . . . , xm }, is deﬁned as range(x) = max(x) − min(x) = x(m) − x(1) .
102
(3.4)
3.2
Summary Statistics
Table 3.4. Range, standard deviation (std), absolute average difference (AAD), median absolute difference (MAD), and interquartile range (IQR) for sepal length, sepal width, petal length, and petal width. (All values are in centimeters.) Measure range std AAD MAD IQR
Sepal Length 3.6 0.8 0.7 0.7 1.3
Sepal Width 2.4 0.4 0.3 0.3 0.5
Petal Length 5.9 1.8 1.6 1.2 3.5
Petal Width 2.4 0.8 0.6 0.7 1.5
Although the range identiﬁes the maximum spread, it can be misleading if most of the values are concentrated in a narrow band of values, but there are also a relatively small number of more extreme values. Hence, the variance is preferred as a measure of spread. The variance of the (observed) values of an attribute x is typically written as s2x and is deﬁned below. The standard deviation, which is the square root of the variance, is written as sx and has the same units as x. 1 = (xi − x)2 m−1 m
variance(x) =
s2x
(3.5)
i=1
The mean can be distorted by outliers, and since the variance is computed using the mean, it is also sensitive to outliers. Indeed, the variance is particularly sensitive to outliers since it uses the squared diﬀerence between the mean and other values. As a result, more robust estimates of the spread of a set of values are often used. Following are the deﬁnitions of three such measures: the absolute average deviation (AAD), the median absolute deviation (MAD), and the interquartile range(IQR). Table 3.4 shows these measures for the Iris data set. 1 xi − x m m
AAD(x) =
(3.6)
i=1
MAD(x) = median {x1 − x, . . . , xm − x}
(3.7)
interquartile range(x) = x75% − x25%
(3.8)
103
Chapter 3
3.2.5
Exploring Data
Multivariate Summary Statistics
Measures of location for data that consists of several attributes (multivariate data) can be obtained by computing the mean or median separately for each attribute. Thus, given a data set the mean of the data objects, x, is given by x = (x1 , . . . , xn ),
(3.9)
where xi is the mean of the ith attribute xi . For multivariate data, the spread of each attribute can be computed independently of the other attributes using any of the approaches described in Section 3.2.4. However, for data with continuous variables, the spread of the data is most commonly captured by the covariance matrix S, whose ij th entry sij is the covariance of the ith and j th attributes of the data. Thus, if xi and xj are the ith and j th attributes, then sij = covariance(xi , xj ).
(3.10)
In turn, covariance(xi , xj ) is given by 1 (xki − xi )(xkj − xj ), covariance(xi , xj ) = m−1 m
(3.11)
k=1
where xki and xkj are the values of the ith and j th attributes for the k th object. Notice that covariance(xi , xi ) = variance(xi ). Thus, the covariance matrix has the variances of the attributes along the diagonal. The covariance of two attributes is a measure of the degree to which two attributes vary together and depends on the magnitudes of the variables. A value near 0 indicates that two attributes do not have a (linear) relationship, but it is not possible to judge the degree of relationship between two variables by looking only at the value of the covariance. Because the correlation of two attributes immediately gives an indication of how strongly two attributes are (linearly) related, correlation is preferred to covariance for data exploration. (Also see the discussion of correlation in Section 2.4.5.) The ij th entry of the correlation matrix R, is the correlation between the ith and j th attributes of the data. If xi and xj are the ith and j th attributes, then rij = correlation(xi , xj ) =
104
covariance(xi , xj ) , si sj
(3.12)
3.3
Visualization
where si and sj are the variances of xi and xj , respectively. The diagonal entries of R are correlation(xi , xi ) = 1, while the other entries are between −1 and 1. It is also useful to consider correlation matrices that contain the pairwise correlations of objects instead of attributes.
3.2.6
Other Ways to Summarize the Data
There are, of course, other types of summary statistics. For instance, the skewness of a set of values measures the degree to which the values are symmetrically distributed around the mean. There are also other characteristics of the data that are not easy to measure quantitatively, such as whether the distribution of values is multimodal; i.e., the data has multiple “bumps” where most of the values are concentrated. In many cases, however, the most eﬀective approach to understanding the more complicated or subtle aspects of how the values of an attribute are distributed, is to view the values graphically in the form of a histogram. (Histograms are discussed in the next section.)
3.3
Visualization
Data visualization is the display of information in a graphic or tabular format. Successful visualization requires that the data (information) be converted into a visual format so that the characteristics of the data and the relationships among data items or attributes can be analyzed or reported. The goal of visualization is the interpretation of the visualized information by a person and the formation of a mental model of the information. In everyday life, visual techniques such as graphs and tables are often the preferred approach used to explain the weather, the economy, and the results of political elections. Likewise, while algorithmic or mathematical approaches are often emphasized in most technical disciplines—data mining included— visual techniques can play a key role in data analysis. In fact, sometimes the use of visualization techniques in data mining is referred to as visual data mining.
3.3.1
Motivations for Visualization
The overriding motivation for using visualization is that people can quickly absorb large amounts of visual information and ﬁnd patterns in it. Consider Figure 3.2, which shows the Sea Surface Temperature (SST) in degrees Celsius for July, 1982. This picture summarizes the information from approximately 250,000 numbers and is readily interpreted in a few seconds. For example, it
105
Chapter 3
Exploring Data
90 30 60 25 30 Latitude
20
0
15
–30
10
5 – 60 0 –90 –180 –150 –120 –90
–60
–30
0 30 Longitude
60
90
120
150 180
Temp
Figure 3.2. Sea Surface Temperature (SST) for July, 1982.
is easy to see that the ocean temperature is highest at the equator and lowest at the poles. Another general motivation for visualization is to make use of the domain knowledge that is “locked up in people’s heads.” While the use of domain knowledge is an important task in data mining, it is often diﬃcult or impossible to fully utilize such knowledge in statistical or algorithmic tools. In some cases, an analysis can be performed using nonvisual tools, and then the results presented visually for evaluation by the domain expert. In other cases, having a domain specialist examine visualizations of the data may be the best way of ﬁnding patterns of interest since, by using domain knowledge, a person can often quickly eliminate many uninteresting patterns and direct the focus to the patterns that are important.
3.3.2
General Concepts
This section explores some of the general concepts related to visualization, in particular, general approaches for visualizing the data and its attributes. A number of visualization techniques are mentioned brieﬂy and will be described in more detail when we discuss speciﬁc approaches later on. We assume that the reader is familiar with line graphs, bar charts, and scatter plots.
106
3.3
Visualization
Representation: Mapping Data to Graphical Elements The ﬁrst step in visualization is the mapping of information to a visual format; i.e., mapping the objects, attributes, and relationships in a set of information to visual objects, attributes, and relationships. That is, data objects, their attributes, and the relationships among data objects are translated into graphical elements such as points, lines, shapes, and colors. Objects are usually represented in one of three ways. First, if only a single categorical attribute of the object is being considered, then objects are often lumped into categories based on the value of that attribute, and these categories are displayed as an entry in a table or an area on a screen. (Examples shown later in this chapter are a crosstabulation table and a bar chart.) Second, if an object has multiple attributes, then the object can be displayed as a row (or column) of a table or as a line on a graph. Finally, an object is often interpreted as a point in two or threedimensional space, where graphically, the point might be represented by a geometric ﬁgure, such as a circle, cross, or box. For attributes, the representation depends on the type of attribute, i.e., nominal, ordinal, or continuous (interval or ratio). Ordinal and continuous attributes can be mapped to continuous, ordered graphical features such as location along the x, y, or z axes; intensity; color; or size (diameter, width, height, etc.). For categorical attributes, each category can be mapped to a distinct position, color, shape, orientation, embellishment, or column in a table. However, for nominal attributes, whose values are unordered, care should be taken when using graphical features, such as color and position that have an inherent ordering associated with their values. In other words, the graphical elements used to represent the ordinal values often have an order, but ordinal values do not. The representation of relationships via graphical elements occurs either explicitly or implicitly. For graph data, the standard graph representation— a set of nodes with links between the nodes—is normally used. If the nodes (data objects) or links (relationships) have attributes or characteristics of their own, then this is represented graphically. To illustrate, if the nodes are cities and the links are highways, then the diameter of the nodes might represent population, while the width of the links might represent the volume of traﬃc. In most cases, though, mapping objects and attributes to graphical elements implicitly maps the relationships in the data to relationships among graphical elements. To illustrate, if the data object represents a physical object that has a location, such as a city, then the relative positions of the graphical objects corresponding to the data objects tend to naturally preserve the actual
107
Chapter 3
Exploring Data
relative positions of the objects. Likewise, if there are two or three continuous attributes that are taken as the coordinates of the data points, then the resulting plot often gives considerable insight into the relationships of the attributes and the data points because data points that are visually close to each other have similar values for their attributes. In general, it is diﬃcult to ensure that a mapping of objects and attributes will result in the relationships being mapped to easily observed relationships among graphical elements. Indeed, this is one of the most challenging aspects of visualization. In any given set of data, there are many implicit relationships, and hence, a key challenge of visualization is to choose a technique that makes the relationships of interest easily observable. Arrangement As discussed earlier, the proper choice of visual representation of objects and attributes is essential for good visualization. The arrangement of items within the visual display is also crucial. We illustrate this with two examples. Example 3.5. This example illustrates the importance of rearranging a table of data. In Table 3.5, which shows nine objects with six binary attributes, there is no clear relationship between objects and attributes, at least at ﬁrst glance. If the rows and columns of this table are permuted, however, as shown in Table 3.6, then it is clear that there are really only two types of objects in the table—one that has all ones for the ﬁrst three attributes and one that has only ones for the last three attributes. Table 3.5. A table of nine objects (rows) with six binary attributes (columns).
1 2 3 4 5 6 7 8 9
108
1 0 1 0 1 0 1 0 1 0
2 1 0 1 0 1 0 1 0 1
3 0 1 0 1 0 1 0 1 0
4 1 0 1 0 1 0 1 0 1
5 1 0 1 0 1 0 1 0 1
6 0 1 0 1 0 1 0 1 0
Table 3.6. A table of nine objects (rows) with six binary attributes (columns) permuted so that the relationships of the rows and columns are clear. 4 2 6 8 5 3 9 1 7
6 1 1 1 1 0 0 0 0 0
1 1 1 1 1 0 0 0 0 0
3 1 1 1 1 0 0 0 0 0
2 0 0 0 0 1 1 1 1 1
5 0 0 0 0 1 1 1 1 1
4 0 0 0 0 1 1 1 1 1
3.3
Visualization
Example 3.6. Consider Figure 3.3(a), which shows a visualization of a graph. If the connected components of the graph are separated, as in Figure 3.3(b), then the relationships between nodes and graphs become much simpler to understand.
(a) Original view of a graph.
(b) Uncoupled view of connected components of the graph.
Figure 3.3. Two visualizations of a graph.
Selection Another key concept in visualization is selection, which is the elimination or the deemphasis of certain objects and attributes. Speciﬁcally, while data objects that only have a few dimensions can often be mapped to a two or threedimensional graphical representation in a straightforward way, there is no completely satisfactory and general approach to represent data with many attributes. Likewise, if there are many data objects, then visualizing all the objects can result in a display that is too crowded. If there are many attributes and many objects, then the situation is even more challenging. The most common approach to handling many attributes is to choose a subset of attributes—usually two—for display. If the dimensionality is not too high, a matrix of bivariate (twoattribute) plots can be constructed for simultaneous viewing. (Figure 3.16 shows a matrix of scatter plots for the pairs of attributes of the Iris data set.) Alternatively, a visualization program can automatically show a series of twodimensional plots, in which the sequence is user directed or based on some predeﬁned strategy. The hope is that visualizing a collection of twodimensional plots will provide a more complete view of the data.
109
Chapter 3
Exploring Data
The technique of selecting a pair (or small number) of attributes is a type of dimensionality reduction, and there are many more sophisticated dimensionality reduction techniques that can be employed, e.g., principal components analysis (PCA). Consult Appendices A (Linear Algebra) and B (Dimensionality Reduction) for more information. When the number of data points is high, e.g., more than a few hundred, or if the range of the data is large, it is diﬃcult to display enough information about each object. Some data points can obscure other data points, or a data object may not occupy enough pixels to allow its features to be clearly displayed. For example, the shape of an object cannot be used to encode a characteristic of that object if there is only one pixel available to display it. In these situations, it is useful to be able to eliminate some of the objects, either by zooming in on a particular region of the data or by taking a sample of the data points.
3.3.3
Techniques
Visualization techniques are often specialized to the type of data being analyzed. Indeed, new visualization techniques and approaches, as well as specialized variations of existing approaches, are being continuously created, typically in response to new kinds of data and visualization tasks. Despite this specialization and the ad hoc nature of visualization, there are some generic ways to classify visualization techniques. One such classiﬁcation is based on the number of attributes involved (1, 2, 3, or many) or whether the data has some special characteristic, such as a hierarchical or graph structure. Visualization methods can also be classiﬁed according to the type of attributes involved. Yet another classiﬁcation is based on the type of application: scientiﬁc, statistical, or information visualization. The following discussion will use three categories: visualization of a small number of attributes, visualization of data with spatial and/or temporal attributes, and visualization of data with many attributes. Most of the visualization techniques discussed here can be found in a wide variety of mathematical and statistical packages, some of which are freely available. There are also a number of data sets that are freely available on the World Wide Web. Readers are encouraged to try these visualization techniques as they proceed through the following sections.
110
3.3
Visualization
Visualizing Small Numbers of Attributes This section examines techniques for visualizing data with respect to a small number of attributes. Some of these techniques, such as histograms, give insight into the distribution of the observed values for a single attribute. Other techniques, such as scatter plots, are intended to display the relationships between the values of two attributes. Stem and Leaf Plots Stem and leaf plots can be used to provide insight into the distribution of onedimensional integer or continuous data. (We will assume integer data initially, and then explain how stem and leaf plots can be applied to continuous data.) For the simplest type of stem and leaf plot, we split the values into groups, where each group contains those values that are the same except for the last digit. Each group becomes a stem, while the last digits of a group are the leaves. Hence, if the values are twodigit integers, e.g., 35, 36, 42, and 51, then the stems will be the highorder digits, e.g., 3, 4, and 5, while the leaves are the loworder digits, e.g., 1, 2, 5, and 6. By plotting the stems vertically and leaves horizontally, we can provide a visual representation of the distribution of the data. Example 3.7. The set of integers shown in Figure 3.4 is the sepal length in centimeters (multiplied by 10 to make the values integers) taken from the Iris data set. For convenience, the values have also been sorted. The stem and leaf plot for this data is shown in Figure 3.5. Each number in Figure 3.4 is ﬁrst put into one of the vertical groups—4, 5, 6, or 7—according to its ten’s digit. Its last digit is then placed to the right of the colon. Often, especially if the amount of data is larger, it is desirable to split the stems. For example, instead of placing all values whose ten’s digit is 4 in the same “bucket,” the stem 4 is repeated twice; all values 40–44 are put in the bucket corresponding to the ﬁrst stem and all values 45–49 are put in the bucket corresponding to the second stem. This approach is shown in the stem and leaf plot of Figure 3.6. Other variations are also possible. Histograms Stem and leaf plots are a type of histogram, a plot that displays the distribution of values for attributes by dividing the possible values into bins and showing the number of objects that fall into each bin. For categorical data, each value is a bin. If this results in too many values, then values are combined in some way. For continuous attributes, the range of values is divided into bins—typically, but not necessarily, of equal width—and the values in each bin are counted.
111
Chapter 3 43 50 54 57 61 65 71
4 5 6 7
44 50 54 57 61 65 72
: : : :
44 50 54 57 61 65 72
44 50 54 57 62 65 72
45 50 54 58 62 65 73
46 50 54 58 62 66 74
Exploring Data 46 46 46 47 47 48 48 48 48 48 49 49 50 50 50 51 51 51 51 51 51 51 51 51 55 55 55 55 55 55 55 56 56 56 56 56 58 58 58 58 58 59 59 59 60 60 60 60 62 63 63 63 63 63 63 63 63 63 64 64 66 67 67 67 67 67 67 67 67 68 68 68 76 77 77 77 77 79 Figure 3.4. Sepal length data from the Iris data set.
49 52 56 60 64 69
49 52 57 60 64 69
49 52 57 61 64 69
49 52 57 61 64 69
50 53 57 61 64 70
34444566667788888999999 0000000000111111111222234444445555555666666777777778888888999 000000111111222233333333344444445555566777777778889999 0122234677779 Figure 3.5. Stem and leaf plot for the sepal length from the Iris data set.
4 4 5 5 6 6 7 7
: : : : : : : :
3444 566667788888999999 000000000011111111122223444444 5555555666666777777778888888999 00000011111122223333333334444444 5555566777777778889999 0122234 677779
Figure 3.6. Stem and leaf plot for the sepal length from the Iris data set when buckets corresponding to digits are split.
Once the counts are available for each bin, a bar plot is constructed such that each bin is represented by one bar and the area of each bar is proportional to the number of values (objects) that fall into the corresponding range. If all intervals are of equal width, then all bars are the same width and the height of a bar is proportional to the number of values in the corresponding bin. Example 3.8. Figure 3.7 shows histograms (with 10 bins) for sepal length, sepal width, petal length, and petal width. Since the shape of a histogram can depend on the number of bins, histograms for the same data, but with 20 bins, are shown in Figure 3.8. There are variations of the histogram plot. A relative (frequency) histogram replaces the count by the relative frequency. However, this is just a
112
3.3
15
40
45
35
40
40
30
35 30 25 20
10
4.5
5
5.5
6
6.5
7
7.5
8
20
5
5
0
0
2
2.5
Sepal Length
3
3.5
4
4.5
25 20 10 5
0
1
2
3
4
5
6
0
7
0
0.5
(b) Sepal width.
1
1.5
2
2.5
3
Petal Width
Petal Length
Sepal Width
(a) Sepal length.
30
15
10
10
4
35
25
15
15
5 0
50 45
Count
20
Count
Count
25
Count
30
Visualization
(c) Petal length.
(d) Petal width.
Figure 3.7. Histograms of four Iris attributes (10 bins).
30
14
25
8 6
20
Count
10
Count
Count
12
15 10
0
5 4
4.5
5
5.5
6
6.5
7
7.5
Sepal Length
(a) Sepal length.
8
0
35
30
30
25
25
20 15 10
4 2
35
Count
16
2.5
3
3.5
4
Sepal Width
(b) Sepal width.
4.5
0
15 10
5 2
20
5
1
2
3
4
5
6
7
0
Petal Length
(c) Petal length.
0
0.5
1
1.5
2
2.5
Petal Width
(d) Petal width.
Figure 3.8. Histograms of four Iris attributes (20 bins).
change in scale of the y axis, and the shape of the histogram does not change. Another common variation, especially for unordered categorical data, is the Pareto histogram, which is the same as a normal histogram except that the categories are sorted by count so that the count is decreasing from left to right. TwoDimensional Histograms Twodimensional histograms are also possible. Each attribute is divided into intervals and the two sets of intervals deﬁne twodimensional rectangles of values. Example 3.9. Figure 3.9 shows a twodimensional histogram of petal length and petal width. Because each attribute is split into three bins, there are nine rectangular twodimensional bins. The height of each rectangular bar indicates the number of objects (ﬂowers in this case) that fall into each bin. Most of the ﬂowers fall into only three of the bins—those along the diagonal. It is not possible to see this by looking at the onedimensional distributions.
113
Chapter 3
Exploring Data
50
Count
40 30 20 10 0 2 1.5
Petal Width
1 0.5
2
3
4
5
6
Petal Length
Figure 3.9. Twodimensional histogram of petal length and width in the Iris data set.
While twodimensional histograms can be used to discover interesting facts about how the values of two attributes cooccur, they are visually more complicated. For instance, it is easy to imagine a situation in which some of the columns are hidden by others. Box Plots Box plots are another method for showing the distribution of the values of a single numerical attribute. Figure 3.10 shows a labeled box plot for sepal length. The lower and upper ends of the box indicate the 25th and 75th percentiles, respectively, while the line inside the box indicates the value of the 50th percentile. The top and bottom lines of the tails indicate the 10th and 90th percentiles. Outliers are shown by “+” marks. Box plots are relatively compact, and thus, many of them can be shown on the same plot. Simpliﬁed versions of the box plot, which take less space, can also be used. Example 3.10. The box plots for the ﬁrst four attributes of the Iris data set are shown in Figure 3.11. Box plots can also be used to compare how attributes vary between diﬀerent classes of objects, as shown in Figure 3.12.
Pie Chart A pie chart is similar to a histogram, but is typically used with categorical attributes that have a relatively small number of values. Instead of showing the relative frequency of diﬀerent values with the area or height of a bar, as in a histogram, a pie chart uses the relative area of a circle to indicate relative frequency. Although pie charts are common in popular articles, they
114
3.3
Visualization
Outlier
+ + +
8
90th percentile
7
Values (centimeters)
6
75th percentile 50th
percentile
25th percentile
5 + + +
4 3 2
+
1
10th percentile
0
+
Sepal Length
Figure 3.10. Description of box plot for sepal length.
+ 2
Values (centimeters)
Values (centimeters)
Values (centimeters)
7
6
3
Petal Width
8
7
5 4
Petal Length
Figure 3.11. Box plot for Iris attributes.
6
1
Sepal Width
5 4 3
+
6 5
+
4 3
2
+
2
+ + 1
0
Sepal Length Sepal Width Petal Length
(a) Setosa.
Petal Width
Sepal Length Sepal Width Petal Length
Petal Width
Sepal Length Sepal Width Petal Length
(b) Versicolour.
Petal Width
(c) Virginica.
Figure 3.12. Box plots of attributes by Iris species.
are used less frequently in technical publications because the size of relative areas can be hard to judge. Histograms are preferred for technical work. Example 3.11. Figure 3.13 displays a pie chart that shows the distribution of Iris species in the Iris data set. In this case, all three ﬂower types have the same frequency. Percentile Plots and Empirical Cumulative Distribution Functions A type of diagram that shows the distribution of the data more quantitatively is the plot of an empirical cumulative distribution function. While this type of plot may sound complicated, the concept is straightforward. For each value of a statistical distribution, a cumulative distribution function (CDF) shows
115
Chapter 3
Exploring Data
Setosa
Virginica
Versicolour
Figure 3.13. Distribution of the types of Iris flowers.
the probability that a point is less than that value. For each observed value, an empirical cumulative distribution function (ECDF) shows the fraction of points that are less than this value. Since the number of points is ﬁnite, the empirical cumulative distribution function is a step function. Example 3.12. Figure 3.14 shows the ECDFs of the Iris attributes. The percentiles of an attribute provide similar information. Figure 3.15 shows the percentile plots of the four continuous attributes of the Iris data set from Table 3.2. The reader should compare these ﬁgures with the histograms given in Figures 3.7 and 3.8. Scatter Plots Most people are familiar with scatter plots to some extent, and they were used in Section 2.4.5 to illustrate linear correlation. Each data object is plotted as a point in the plane using the values of the two attributes as x and y coordinates. It is assumed that the attributes are either integer or realvalued. Example 3.13. Figure 3.16 shows a scatter plot for each pair of attributes of the Iris data set. The diﬀerent species of Iris are indicated by diﬀerent markers. The arrangement of the scatter plots of pairs of attributes in this type of tabular format, which is known as a scatter plot matrix, provides an organized way to examine a number of scatter plots simultaneously.
116
1
1
0.9
0.9
0.8
0.8
0.7
0.7
0.6
0.6 F(x)
F(x)
3.3
0.5
0.5
0.4
0.4
0.3
0.3
0.2
0.2
0.1 0 4
0.1
4.5
5
5.5
6 x
6.5
7
7.5
0 2
8
2.5
0.8
0.8
0.7
0.7
0.6
0.6 F(x)
1 0.9
0.5
4
4.5
2
2.5
0.5
0.4
0.4
0.3
0.3
0.2
0.2
0.1
0.1
4 x
5
6
0 0
7
0.5
1
1.5 x
(c) Petal Length.
(d) Petal Width.
Figure 3.14. Empirical CDFs of four Iris attributes.
sepal length sepal width petal length petal width
7 6 Value (centimeters)
F(x)
1
3
3.5
(b) Sepal Width.
0.9
2
3 x
(a) Sepal Length.
0 1
Visualization
5 4 3 2 1
0
20
40
60
80
10
Percentile
Figure 3.15. Percentile plots for sepal length, sepal width, petal length, and petal width.
117
sepal length
sepal width
petal length
118
petal width 0
0.5
1
1.5
2
2.5
2
4
6
2
2.5
3
3.5
4
4.5
5
6
7
8
6
7
sepal length
5
Virginica
Versicolour
Setosa
3
sepal width
4
2
4
petal length
6
Figure 3.16. Matrix of scatter plots for the Iris data set.
8 2
0
1
petal width
2
Chapter 3 Exploring Data
3.3
Visualization
There are two main uses for scatter plots. First, they graphically show the relationship between two attributes. In Section 2.4.5, we saw how scatter plots could be used to judge the degree of linear correlation. (See Figure 2.17.) Scatter plots can also be used to detect nonlinear relationships, either directly or by using a scatter plot of the transformed attributes. Second, when class labels are available, they can be used to investigate the degree to which two attributes separate the classes. If is possible to draw a line (or a more complicated curve) that divides the plane deﬁned by the two attributes into separate regions that contain mostly objects of one class, then it is possible to construct an accurate classiﬁer based on the speciﬁed pair of attributes. If not, then more attributes or more sophisticated methods are needed to build a classiﬁer. In Figure 3.16, many of the pairs of attributes (for example, petal width and petal length) provide a moderate separation of the Iris species. Example 3.14. There are two separate approaches for displaying three attributes of a data set with a scatter plot. First, each object can be displayed according to the values of three, instead of two attributes. Figure 3.17 shows a threedimensional scatter plot for three attributes in the Iris data set. Second, one of the attributes can be associated with some characteristic of the marker, such as its size, color, or shape. Figure 3.18 shows a plot of three attributes of the Iris data set, where one of the attributes, sepal width, is mapped to the size of the marker. Extending Two and ThreeDimensional Plots As illustrated by Figure 3.18, two or threedimensional plots can be extended to represent a few additional attributes. For example, scatter plots can display up to three additional attributes using color or shading, size, and shape, allowing ﬁve or six dimensions to be represented. There is a need for caution, however. As the complexity of a visual representation of the data increases, it becomes harder for the intended audience to interpret the information. There is no beneﬁt in packing six dimensions’ worth of information into a two or threedimensional plot, if doing so makes it impossible to understand. Visualizing Spatiotemporal Data Data often has spatial or temporal attributes. For instance, the data may consist of a set of observations on a spatial grid, such as observations of pressure on the surface of the Earth or the modeled temperature at various grid points in the simulation of a physical object. These observations can also be
119
Chapter 3
Exploring Data
Setosa Versicolour Virginica
Sepal Length
2 1.5 1 0.5 0 5 4 Sepal Width
1
3
2
3
2
7
6
5
4 Petal Width
Figure 3.17. Threedimensional scatter plot of sepal width, sepal length, and petal width.
2.5
Setosa Versicolour Virginica
Petal Width
2
1.5
1
0.5
0 1
2
3
4 Petal Length
5
6
7
Figure 3.18. Scatter plot of petal length versus petal width, with the size of the marker indicating sepal width.
120
3.3
25 5
0
5 5
0
0
5
50
5 0
5
5
Visualization
5
10
10
25
25
15
20
20
20
15
5
15
20
0 5 1 20
25
15
25
10
25
25
25
15
5
25
25 20
20
10
20
15
0
5
10
5
5
5
5
5
–5 5
Temperature (Celsius)
Figure 3.19. Contour plot of SST for December 1998.
made at various points in time. In addition, data may have only a temporal component, such as time series data that gives the daily prices of stocks. Contour Plots For some threedimensional data, two attributes specify a position in a plane, while the third has a continuous value, such as temperature or elevation. A useful visualization for such data is a contour plot, which breaks the plane into separate regions where the values of the third attribute (temperature, elevation) are roughly the same. A common example of a contour plot is a contour map that shows the elevation of land locations. Example 3.15. Figure 3.19 shows a contour plot of the average sea surface temperature (SST) for December 1998. The land is arbitrarily set to have a temperature of 0◦ C. In many contour maps, such as that of Figure 3.19, the contour lines that separate two regions are labeled with the value used to separate the regions. For clarity, some of these labels have been deleted. Surface Plots Like contour plots, surface plots use two attributes for the x and y coordinates. The third attribute is used to indicate the height above
121
Chapter 3
Exploring Data
(a) Set of 12 points.
(b) Overall density function—surface plot.
Figure 3.20. Density of a set of 12 points.
the plane deﬁned by the ﬁrst two attributes. While such graphs can be useful, they require that a value of the third attribute be deﬁned for all combinations of values for the ﬁrst two attributes, at least over some range. Also, if the surface is too irregular, then it can be diﬃcult to see all the information, unless the plot is viewed interactively. Thus, surface plots are often used to describe mathematical functions or physical surfaces that vary in a relatively smooth manner. Example 3.16. Figure 3.20 shows a surface plot of the density around a set of 12 points. This example is further discussed in Section 9.3.3. Vector Field Plots In some data, a characteristic may have both a magnitude and a direction associated with it. For example, consider the ﬂow of a substance or the change of density with location. In these situations, it can be useful to have a plot that displays both direction and magnitude. This type of plot is known as a vector plot. Example 3.17. Figure 3.21 shows a contour plot of the density of the two smaller density peaks from Figure 3.20(b), annotated with the density gradient vectors. LowerDimensional Slices Consider a spatiotemporal data set that records some quantity, such as temperature or pressure, at various locations over time. Such a data set has four dimensions and cannot be easily displayed by the types
122
3.3
Visualization
Figure 3.21. Vector plot of the gradient (change) in density for the bottom two density peaks of Figure 3.20.
of plots that we have described so far. However, separate “slices” of the data can be displayed by showing a set of plots, one for each month. By examining the change in a particular area from one month to another, it is possible to notice changes that occur, including those that may be due to seasonal factors. Example 3.18. The underlying data set for this example consists of the average monthly sea level pressure (SLP) from 1982 to 1999 on a 2.5◦ by 2.5◦ latitudelongitude grid. The twelve monthly plots of pressure for one year are shown in Figure 3.22. In this example, we are interested in slices for a particular month in the year 1982. More generally, we can consider slices of the data along any arbitrary dimension. Animation Another approach to dealing with slices of data, whether or not time is involved, is to employ animation. The idea is to display successive twodimensional slices of the data. The human visual system is well suited to detecting visual changes and can often notice changes that might be diﬃcult to detect in another manner. Despite the visual appeal of animation, a set of still plots, such as those of Figure 3.22, can be more useful since this type of visualization allows the information to be studied in arbitrary order and for arbitrary amounts of time.
123
Chapter 3
Exploring Data
January
February
March
April
May
June
July
August
September
October
November
December
Figure 3.22. Monthly plots of sea level pressure over the 12 months of 1982.
3.3.4
Visualizing HigherDimensional Data
This section considers visualization techniques that can display more than the handful of dimensions that can be observed with the techniques just discussed. However, even these techniques are somewhat limited in that they only show some aspects of the data. Matrices An image can be regarded as a rectangular array of pixels, where each pixel is characterized by its color and brightness. A data matrix is a rectangular array of values. Thus, a data matrix can be visualized as an image by associating each entry of the data matrix with a pixel in the image. The brightness or color of the pixel is determined by the value of the corresponding entry of the matrix.
124
3.3
Visualization
3
2 1.5
Setosa
Setosa
2.5
0.9
50
0 0.5
100
Versicolour
Versicolour
0.5
0.8
50
1
0.7
0.6
100 1.5 2 150
Sepal length
Sepal width
Petal length
Petal width
Standard Deviation
0.5
Virginica
Virginica
1
0.4 150
50 Setosa
Figure 3.23. Plot of the Iris data matrix where columns have been standardized to have a mean of 0 and standard deviation of 1.
100 Versicolour
Correlation Virginica
Figure 3.24. Plot of the Iris correlation matrix.
There are some important practical considerations when visualizing a data matrix. If class labels are known, then it is useful to reorder the data matrix so that all objects of a class are together. This makes it easier, for example, to detect if all objects in a class have similar attribute values for some attributes. If diﬀerent attributes have diﬀerent ranges, then the attributes are often standardized to have a mean of zero and a standard deviation of 1. This prevents the attribute with the largest magnitude values from visually dominating the plot. Example 3.19. Figure 3.23 shows the standardized data matrix for the Iris data set. The ﬁrst 50 rows represent Iris ﬂowers of the species Setosa, the next 50 Versicolour, and the last 50 Virginica. The Setosa ﬂowers have petal width and length well below the average, while the Versicolour ﬂowers have petal width and length around average. The Virginica ﬂowers have petal width and length above average. It can also be useful to look for structure in the plot of a proximity matrix for a set of data objects. Again, it is useful to sort the rows and columns of the similarity matrix (when class labels are known) so that all the objects of a class are together. This allows a visual evaluation of the cohesiveness of each class and its separation from other classes. Example 3.20. Figure 3.24 shows the correlation matrix for the Iris data set. Again, the rows and columns are organized so that all the ﬂowers of a particular species are together. The ﬂowers in each group are most similar
125
Chapter 3
Exploring Data
to each other, but Versicolour and Virginica are more similar to one another than to Setosa. If class labels are not known, various techniques (matrix reordering and seriation) can be used to rearrange the rows and columns of the similarity matrix so that groups of highly similar objects and attributes are together and can be visually identiﬁed. Eﬀectively, this is a simple kind of clustering. See Section 8.5.3 for a discussion of how a proximity matrix can be used to investigate the cluster structure of data. Parallel Coordinates Parallel coordinates have one coordinate axis for each attribute, but the diﬀerent axes are parallel to one other instead of perpendicular, as is traditional. Furthermore, an object is represented as a line instead of as a point. Speciﬁcally, the value of each attribute of an object is mapped to a point on the coordinate axis associated with that attribute, and these points are then connected to form the line that represents the object. It might be feared that this would yield quite a mess. However, in many cases, objects tend to fall into a small number of groups, where the points in each group have similar values for their attributes. If so, and if the number of data objects is not too large, then the resulting parallel coordinates plot can reveal interesting patterns. Example 3.21. Figure 3.25 shows a parallel coordinates plot of the four numerical attributes of the Iris data set. The lines representing objects of diﬀerent classes are distinguished by their shading and the use of three diﬀerent line styles—solid, dotted, and dashed. The parallel coordinates plot shows that the classes are reasonably well separated for petal width and petal length, but less well separated for sepal length and sepal width. Figure 3.25 is another parallel coordinates plot of the same data, but with a diﬀerent ordering of the axes. One of the drawbacks of parallel coordinates is that the detection of patterns in such a plot may depend on the order. For instance, if lines cross a lot, the picture can become confusing, and thus, it can be desirable to order the coordinate axes to obtain sequences of axes with less crossover. Compare Figure 3.26, where sepal width (the attribute that is most mixed) is at the left of the ﬁgure, to Figure 3.25, where this attribute is in the middle. Star Coordinates and Chernoﬀ Faces Another approach to displaying multidimensional data is to encode objects as glyphs or icons—symbols that impart information nonverbally. More
126
3.3
Visualization
8 Setosa Versicolour Virginica
7
Value (centimeters)
6 5 4 3 2 1 0 Sepal Length
Sepal Width
Petal Length
Petal Width
Figure 3.25. A parallel coordinates plot of the four Iris attributes. 8 Setosa Versicolour Virginica
7
Value (centimeters)
6 5 4 3 2 1 0 Sepal Width
Sepal Length
Petal Length
Petal Width
Figure 3.26. A parallel coordinates plot of the four Iris attributes with the attributes reordered to emphasize similarities and dissimilarities of groups.
127
Chapter 3
Exploring Data
speciﬁcally, each attribute of an object is mapped to a particular feature of a glyph, so that the value of the attribute determines the exact nature of the feature. Thus, at a glance, we can distinguish how two objects diﬀer. Star coordinates are one example of this approach. This technique uses one axis for each attribute. These axes all radiate from a center point, like the spokes of a wheel, and are evenly spaced. Typically, all the attribute values are mapped to the range [0,1]. An object is mapped onto this starshaped set of axes using the following process: Each attribute value of the object is converted to a fraction that represents its distance between the minimum and maximum values of the attribute. This fraction is mapped to a point on the axis corresponding to this attribute. Each point is connected with a line segment to the point on the axis preceding or following its own axis; this forms a polygon. The size and shape of this polygon gives a visual description of the attribute values of the object. For ease of interpretation, a separate set of axes is used for each object. In other words, each object is mapped to a polygon. An example of a star coordinates plot of ﬂower 150 is given in Figure 3.27(a). It is also possible to map the values of features to those of more familiar objects, such as faces. This technique is named Chernoﬀ faces for its creator, Herman Chernoﬀ. In this technique, each attribute is associated with a speciﬁc feature of a face, and the attribute value is used to determine the way that the facial feature is expressed. Thus, the shape of the face may become more elongated as the value of the corresponding data feature increases. An example of a Chernoﬀ face for ﬂower 150 is given in Figure 3.27(b). The program that we used to make this face mapped the features to the four features listed below. Other features of the face, such as width between the eyes and length of the mouth, are given default values. Data Feature sepal length sepal width petal length petal width
Facial Feature size of face forehead/jaw relative arc length shape of forehead shape of jaw
Example 3.22. A more extensive illustration of these two approaches to viewing multidimensional data is provided by Figures 3.28 and 3.29, which shows the star and face plots, respectively, of 15 ﬂowers from the Iris data set. The ﬁrst 5 ﬂowers are of species Setosa, the second 5 are Versicolour, and the last 5 are Virginica.
128
Visualization
sepal width
3.3
sepal length petal width
petal length
(a) Star graph of Iris 150.
(b) Chernoﬀ face of Iris 150.
Figure 3.27. Star coordinates graph and Chernoff face of the 150th flower of the Iris data set.
1
2
3
4
5
51
52
53
54
55
101
102
103
104
105
Figure 3.28. Plot of 15 Iris flowers using star coordinates.
1
2
3
4
5
51
52
53
54
55
101
102
103
104
105
Figure 3.29. A plot of 15 Iris flowers using Chernoff faces.
129
Chapter 3
Exploring Data
Despite the visual appeal of these sorts of diagrams, they do not scale well, and thus, they are of limited use for many data mining problems. Nonetheless, they may still be of use as a means to quickly compare small sets of objects that have been selected by other techniques.
3.3.5
Do’s and Don’ts
To conclude this section on visualization, we provide a short list of visualization do’s and don’ts. While these guidelines incorporate a lot of visualization wisdom, they should not be followed blindly. As always, guidelines are no substitute for thoughtful consideration of the problem at hand. ACCENT Principles The following are the ACCENT principles for effective graphical display put forth by D. A. Burn (as adapted by Michael Friendly): Apprehension Ability to correctly perceive relations among variables. Does the graph maximize apprehension of the relations among variables? Clarity Ability to visually distinguish all the elements of a graph. Are the most important elements or relations visually most prominent? Consistency Ability to interpret a graph based on similarity to previous graphs. Are the elements, symbol shapes, and colors consistent with their use in previous graphs? Eﬃciency Ability to portray a possibly complex relation in as simple a way as possible. Are the elements of the graph economically used? Is the graph easy to interpret? Necessity The need for the graph, and the graphical elements. Is the graph a more useful way to represent the data than alternatives (table, text)? Are all the graph elements necessary to convey the relations? Truthfulness Ability to determine the true value represented by any graphical element by its magnitude relative to the implicit or explicit scale. Are the graph elements accurately positioned and scaled? Tufte’s Guidelines Edward R. Tufte has also enumerated the following principles for graphical excellence:
130
3.4
OLAP and Multidimensional Data Analysis
• Graphical excellence is the welldesigned presentation of interesting data— a matter of substance, of statistics, and of design. • Graphical excellence consists of complex ideas communicated with clarity, precision, and eﬃciency. • Graphical excellence is that which gives to the viewer the greatest number of ideas in the shortest time with the least ink in the smallest space. • Graphical excellence is nearly always multivariate. • And graphical excellence requires telling the truth about the data.
3.4
OLAP and Multidimensional Data Analysis
In this section, we investigate the techniques and insights that come from viewing data sets as multidimensional arrays. A number of database systems support such a viewpoint, most notably, OnLine Analytical Processing (OLAP) systems. Indeed, some of the terminology and capabilities of OLAP systems have made their way into spreadsheet programs that are used by millions of people. OLAP systems also have a strong focus on the interactive analysis of data and typically provide extensive capabilities for visualizing the data and generating summary statistics. For these reasons, our approach to multidimensional data analysis will be based on the terminology and concepts common to OLAP systems.
3.4.1
Representing Iris Data as a Multidimensional Array
Most data sets can be represented as a table, where each row is an object and each column is an attribute. In many cases, it is also possible to view the data as a multidimensional array. We illustrate this approach by representing the Iris data set as a multidimensional array. Table 3.7 was created by discretizing the petal length and petal width attributes to have values of low, medium, and high and then counting the number of ﬂowers from the Iris data set that have particular combinations of petal width, petal length, and species type. (For petal width, the categories low, medium, and high correspond to the intervals [0, 0.75), [0.75, 1.75), [1.75, ∞), respectively. For petal length, the categories low, medium, and high correspond to the intervals [0, 2.5), [2.5, 5), [5, ∞), respectively.)
131
Chapter 3
Exploring Data
Table 3.7. Number of flowers having a particular combination of petal width, petal length, and species type. Petal Length low low medium medium medium medium high high high high
Petal Width low medium low medium high high medium medium high high
Species Type Setosa Setosa Setosa Versicolour Versicolour Virginica Versicolour Virginica Versicolour Virginica
Count 46 2 2 43 3 3 2 3 2 44
Petal Width
Virginica Versicolour Setosa high
0
0
0
medium
0
0
2
2
46 low
Petal Width
0
medium
low
high
es
i ec
Sp
Figure 3.30. A multidimensional data representation for the Iris data set.
132
OLAP and Multidimensional Data Analysis
Length
Table 3.8. Crosstabulation of flowers according to petal length and width for flowers of the Setosa species.
low medium high
low 46 2 0
Width medium 2 0 0
high 0 0 0
Table 3.9. Crosstabulation of flowers according to petal length and width for flowers of the Versicolour species.
Length
3.4
low medium high
low 0 0 0
Width medium 0 43 2
high 0 3 2
Length
Table 3.10. Crosstabulation of flowers according to petal length and width for flowers of the Virginica species.
low medium high
low 0 0 0
Width medium 0 0 3
high 0 3 44
Empty combinations—those combinations that do not correspond to at least one ﬂower—are not shown. The data can be organized as a multidimensional array with three dimensions corresponding to petal width, petal length, and species type, as illustrated in Figure 3.30. For clarity, slices of this array are shown as a set of three twodimensional tables, one for each species—see Tables 3.8, 3.9, and 3.10. The information contained in both Table 3.7 and Figure 3.30 is the same. However, in the multidimensional representation shown in Figure 3.30 (and Tables 3.8, 3.9, and 3.10), the values of the attributes—petal width, petal length, and species type—are array indices. What is important are the insights can be gained by looking at data from a multidimensional viewpoint. Tables 3.8, 3.9, and 3.10 show that each species of Iris is characterized by a diﬀerent combination of values of petal length and width. Setosa ﬂowers have low width and length, Versicolour ﬂowers have medium width and length, and Virginica ﬂowers have high width and length.
3.4.2
Multidimensional Data: The General Case
The previous section gave a speciﬁc example of using a multidimensional approach to represent and analyze a familiar data set. Here we describe the general approach in more detail.
133
Chapter 3
Exploring Data
The starting point is usually a tabular representation of the data, such as that of Table 3.7, which is called a fact table. Two steps are necessary in order to represent data as a multidimensional array: identiﬁcation of the dimensions and identiﬁcation of an attribute that is the focus of the analysis. The dimensions are categorical attributes or, as in the previous example, continuous attributes that have been converted to categorical attributes. The values of an attribute serve as indices into the array for the dimension corresponding to the attribute, and the number of attribute values is the size of that dimension. In the previous example, each attribute had three possible values, and thus, each dimension was of size three and could be indexed by three values. This produced a 3 × 3 × 3 multidimensional array. Each combination of attribute values (one value for each diﬀerent attribute) deﬁnes a cell of the multidimensional array. To illustrate using the previous example, if petal length = low, petal width = medium, and species = Setosa, a speciﬁc cell containing the value 2 is identiﬁed. That is, there are only two ﬂowers in the data set that have the speciﬁed attribute values. Notice that each row (object) of the data set in Table 3.7 corresponds to a cell in the multidimensional array. The contents of each cell represents the value of a target quantity (target variable or attribute) that we are interested in analyzing. In the Iris example, the target quantity is the number of ﬂowers whose petal width and length fall within certain limits. The target attribute is quantitative because a key goal of multidimensional data analysis is to look aggregate quantities, such as totals or averages. The following summarizes the procedure for creating a multidimensional data representation from a data set represented in tabular form. First, identify the categorical attributes to be used as the dimensions and a quantitative attribute to be used as the target of the analysis. Each row (object) in the table is mapped to a cell of the multidimensional array. The indices of the cell are speciﬁed by the values of the attributes that were selected as dimensions, while the value of the cell is the value of the target attribute. Cells not deﬁned by the data are assumed to have a value of 0. Example 3.23. To further illustrate the ideas just discussed, we present a more traditional example involving the sale of products.The fact table for this example is given by Table 3.11. The dimensions of the multidimensional representation are the product ID, location, and date attributes, while the target attribute is the revenue. Figure 3.31 shows the multidimensional representation of this data set. This larger and more complicated data set will be used to illustrate additional concepts of multidimensional data analysis.
134
3.4
3.4.3
OLAP and Multidimensional Data Analysis
Analyzing Multidimensional Data
In this section, we describe diﬀerent multidimensional analysis techniques. In particular, we discuss the creation of data cubes, and related operations, such as slicing, dicing, dimensionality reduction, rollup, and drill down. Data Cubes: Computing Aggregate Quantities A key motivation for taking a multidimensional viewpoint of data is the importance of aggregating data in various ways. In the sales example, we might wish to ﬁnd the total sales revenue for a speciﬁc year and a speciﬁc product. Or we might wish to see the yearly sales revenue for each location across all products. Computing aggregate totals involves ﬁxing speciﬁc values for some of the attributes that are being used as dimensions and then summing over all possible values for the attributes that make up the remaining dimensions. There are other types of aggregate quantities that are also of interest, but for simplicity, this discussion will use totals (sums). Table 3.12 shows the result of summing over all locations for various combinations of date and product. For simplicity, assume that all the dates are within one year. If there are 365 days in a year and 1000 products, then Table 3.12 has 365,000 entries (totals), one for each productdata pair. We could also specify the store location and date and sum over products, or specify the location and product and sum over all dates. Table 3.13 shows the marginal totals of Table 3.12. These totals are the result of further summing over either dates or products. In Table 3.13, the total sales revenue due to product 1, which is obtained by summing across row 1 (over all dates), is $370,000. The total sales revenue on January 1, 2004, which is obtained by summing down column 1 (over all products), is $527,362. The total sales revenue, which is obtained by summing over all rows and columns (all times and products) is $227,352,127. All of these totals are for all locations because the entries of Table 3.13 include all locations. A key point of this example is that there are a number of diﬀerent totals (aggregates) that can be computed for a multidimensional array, depending on how many attributes we sum over. Assume that there are n dimensions and that the ith dimension (attribute) has si possible values. There are n diﬀerent ways to sum only over a single attribute. If we sum over dimension j, then we obtain s1 ∗ · · · ∗ sj−1 ∗ sj+1 ∗ · · · ∗ sn totals, one for each possible combination of attribute values of the n − 1 other attributes (dimensions). The totals that result from summing over one attribute form a multidimensional array of n − 1 dimensions and there are n such arrays of totals. In the sales example, there
135
Chapter 3
Exploring Data
Table 3.11. Sales revenue of products (in dollars) for various locations and times. Product ID .. . 1 1 .. .
Location .. . Minneapolis Chicago .. .
Date .. . Oct. 18, 2004 Oct. 18, 2004 .. .
Revenue .. . $250 $79
1 .. .
Paris .. .
Oct. 18, 2004 .. .
301 .. .
27 27 .. .
Minneapolis Chicago .. .
Oct. 18, 2004 Oct. 18, 2004 .. .
$2,321 $3,278
27 .. .
Paris .. .
Oct. 18, 2004 .. .
$1,325 .. .
$
$
ca
$
Lo
...
tio
n
..
.
...
Date
Product ID Figure 3.31. Multidimensional data representation for sales data.
136
3.4
OLAP and Multidimensional Data Analysis
product ID
Table 3.12. Totals that result from summing over all locations for a fixed time and product.
1 .. . 27 .. .
Jan 1, 2004 $1,001 .. . $10,265 .. .
date Jan 2, 2004 $987
... ...
Dec 31, 2004 $891 .. .
$10,225
...
$9,325 .. .
product ID
Table 3.13. Table 3.12 with marginal totals.
1 .. . 27 .. . total
date Jan 2, 2004 $987
... ...
Dec 31, 2004 $891 .. .
$10,265 .. .
$10,225
...
$9,325 .. .
$527,362
$532,953
...
$631,221
Jan 1, 2004 $1,001 .. .
total $370,000 .. . $3,800,020 .. . $227,352,127
are three sets of totals that result from summing over only one dimension and each set of totals can be displayed as a twodimensional table. If we sum over two dimensions (perhaps starting with one of the arrays of totals obtained by summing over one dimension), then we will obtain a multidimensional array of totals with n − 2 dimensions. There will be n2 distinct arrays of such totals. For the sales examples, there will be 32 = 3 arrays of totals that result from summing over location and product, location and n time, or product and time. In general, summing over k dimensions yields k arrays of totals, each with dimension n − k. A multidimensional representation of the data, together with all possible totals (aggregates), is known as a data cube. Despite the name, the size of each dimension—the number of attribute values—does not need to be equal. Also, a data cube may have either more or fewer than three dimensions. More importantly, a data cube is a generalization of what is known in statistical terminology as a crosstabulation. If marginal totals were added, Tables 3.8, 3.9, or 3.10 would be typical examples of cross tabulations.
137
Chapter 3
Exploring Data
Dimensionality Reduction and Pivoting The aggregation described in the last section can be viewed as a form of dimensionality reduction. Speciﬁcally, the j th dimension is eliminated by summing over it. Conceptually, this collapses each “column” of cells in the j th dimension into a single cell. For both the sales and Iris examples, aggregating over one dimension reduces the dimensionality of the data from 3 to 2. If sj is the number of possible values of the j th dimension, the number of cells is reduced by a factor of sj . Exercise 17 on page 143 asks the reader to explore the diﬀerence between this type of dimensionality reduction and that of PCA. Pivoting refers to aggregating over all dimensions except two. The result is a twodimensional cross tabulation with the two speciﬁed dimensions as the only remaining dimensions. Table 3.13 is an example of pivoting on date and product. Slicing and Dicing These two colorful names refer to rather straightforward operations. Slicing is selecting a group of cells from the entire multidimensional array by specifying a speciﬁc value for one or more dimensions. Tables 3.8, 3.9, and 3.10 are three slices from the Iris set that were obtained by specifying three separate values for the species dimension. Dicing involves selecting a subset of cells by specifying a range of attribute values. This is equivalent to deﬁning a subarray from the complete array. In practice, both operations can also be accompanied by aggregation over some dimensions. RollUp and DrillDown In Chapter 2, attribute values were regarded as being “atomic” in some sense. However, this is not always the case. In particular, each date has a number of properties associated with it such as the year, month, and week. The data can also be identiﬁed as belonging to a particular business quarter, or if the application relates to education, a school quarter or semester. A location also has various properties: continent, country, state (province, etc.), and city. Products can also be divided into various categories, such as clothing, electronics, and furniture. Often these categories can be organized as a hierarchical tree or lattice. For instance, years consist of months or weeks, both of which consist of days. Locations can be divided into nations, which contain states (or other units of local government), which in turn contain cities. Likewise, any category
138
3.5
Bibliographic Notes
of products can be further subdivided. For example, the product category, furniture, can be subdivided into the subcategories, chairs, tables, sofas, etc. This hierarchical structure gives rise to the rollup and drilldown operations. To illustrate, starting with the original sales data, which is a multidimensional array with entries for each date, we can aggregate (roll up) the sales across all the dates in a month. Conversely, given a representation of the data where the time dimension is broken into months, we might want to split the monthly sales totals (drill down) into daily sales totals. Of course, this requires that the underlying sales data be available at a daily granularity. Thus, rollup and drilldown operations are related to aggregation. Notice, however, that they diﬀer from the aggregation operations discussed until now in that they aggregate cells within a dimension, not across the entire dimension.
3.4.4
Final Comments on Multidimensional Data Analysis
Multidimensional data analysis, in the sense implied by OLAP and related systems, consists of viewing the data as a multidimensional array and aggregating data in order to better analyze the structure of the data. For the Iris data, the diﬀerences in petal width and length are clearly shown by such an analysis. The analysis of business data, such as sales data, can also reveal many interesting patterns, such as proﬁtable (or unproﬁtable) stores or products. As mentioned, there are various types of database systems that support the analysis of multidimensional data. Some of these systems are based on relational databases and are known as ROLAP systems. More specialized database systems that speciﬁcally employ a multidimensional data representation as their fundamental data model have also been designed. Such systems are known as MOLAP systems. In addition to these types of systems, statistical databases (SDBs) have been developed to store and analyze various types of statistical data, e.g., census and public health data, that are collected by governments or other large organizations. References to OLAP and SDBs are provided in the bibliographic notes.
3.5
Bibliographic Notes
Summary statistics are discussed in detail in most introductory statistics books, such as [92]. References for exploratory data analysis are the classic text by Tukey [104] and the book by Velleman and Hoaglin [105]. The basic visualization techniques are readily available, being an integral part of most spreadsheets (Microsoft EXCEL [95]), statistics programs (SAS
139
Chapter 3
Exploring Data
[99], SPSS [102], R [96], and SPLUS [98]), and mathematics software (MATLAB [94] and Mathematica [93]). Most of the graphics in this chapter were generated using MATLAB. The statistics package R is freely available as an open source software package from the R project. The literature on visualization is extensive, covering many ﬁelds and many decades. One of the classics of the ﬁeld is the book by Tufte [103]. The book by Spence [101], which strongly inﬂuenced the visualization portion of this chapter, is a useful reference for information visualization—both principles and techniques. This book also provides a thorough discussion of many dynamic visualization techniques that were not covered in this chapter. Two other books on visualization that may also be of interest are those by Card et al. [87] and Fayyad et al. [89]. Finally, there is a great deal of information available about data visualization on the World Wide Web. Since Web sites come and go frequently, the best strategy is a search using “information visualization,” “data visualization,” or “statistical graphics.” However, we do want to single out for attention “The Gallery of Data Visualization,” by Friendly [90]. The ACCENT Principles for eﬀective graphical display as stated in this chapter can be found there, or as originally presented in the article by Burn [86]. There are a variety of graphical techniques that can be used to explore whether the distribution of the data is Gaussian or some other speciﬁed distribution. Also, there are plots that display whether the observed values are statistically signiﬁcant in some sense. We have not covered any of these techniques here and refer the reader to the previously mentioned statistical and mathematical packages. Multidimensional analysis has been around in a variety of forms for some time. One of the original papers was a white paper by Codd [88], the father of relational databases. The data cube was introduced by Gray et al. [91], who described various operations for creating and manipulating data cubes within a relational database framework. A comparison of statistical databases and OLAP is given by Shoshani [100]. Speciﬁc information on OLAP can be found in documentation from database vendors and many popular books. Many database textbooks also have general discussions of OLAP, often in the context of data warehousing. For example, see the text by Ramakrishnan and Gehrke [97].
Bibliography [86] D. A. Burn. Designing Eﬀective Statistical Graphs. In C. R. Rao, editor, Handbook of Statistics 9. Elsevier/NorthHolland, Amsterdam, The Netherlands, September 1993.
140
3.6
Exercises
[87] S. K. Card, J. D. MacKinlay, and B. Shneiderman, editors. Readings in Information Visualization: Using Vision to Think. Morgan Kaufmann Publishers, San Francisco, CA, January 1999. [88] E. F. Codd, S. B. Codd, and C. T. Smalley. Providing OLAP (Online Analytical Processing) to User Analysts: An IT Mandate. White Paper, E.F. Codd and Associates, 1993. [89] U. M. Fayyad, G. G. Grinstein, and A. Wierse, editors. Information Visualization in Data Mining and Knowledge Discovery. Morgan Kaufmann Publishers, San Francisco, CA, September 2001. [90] M. Friendly. Gallery of Data Visualization. http://www.math.yorku.ca/SCS/Gallery/, 2005. [91] J. Gray, S. Chaudhuri, A. Bosworth, A. Layman, D. Reichart, M. Venkatrao, F. Pellow, and H. Pirahesh. Data Cube: A Relational Aggregation Operator Generalizing GroupBy, CrossTab, and SubTotals. Journal Data Mining and Knowledge Discovery, 1(1): 29–53, 1997. [92] B. W. Lindgren. Statistical Theory. CRC Press, January 1993. [93] Mathematica 5.1. Wolfram Research, Inc. http://www.wolfram.com/, 2005. [94] MATLAB 7.0. The MathWorks, Inc. http://www.mathworks.com, 2005. [95] Microsoft Excel 2003. Microsoft, Inc. http://www.microsoft.com/, 2003. [96] R: A language and environment for statistical computing and graphics. The R Project for Statistical Computing. http://www.rproject.org/, 2005. [97] R. Ramakrishnan and J. Gehrke. Database Management Systems. McGrawHill, 3rd edition, August 2002. [98] SPLUS. Insightful Corporation. http://www.insightful.com, 2005. [99] SAS: Statistical Analysis System. SAS Institute Inc. http://www.sas.com/, 2005. [100] A. Shoshani. OLAP and statistical databases: similarities and diﬀerences. In Proc. of the Sixteenth ACM SIGACTSIGMODSIGART Symp. on Principles of Database Systems, pages 185–196. ACM Press, 1997. [101] R. Spence. Information Visualization. ACM Press, New York, December 2000. [102] SPSS: Statistical Package for the Social Sciences. SPSS, Inc. http://www.spss.com/, 2005. [103] E. R. Tufte. The Visual Display of Quantitative Information. Graphics Press, Cheshire, CT, March 1986. [104] J. W. Tukey. Exploratory data analysis. AddisonWesley, 1977. [105] P. Velleman and D. Hoaglin. The ABC’s of EDA: Applications, Basics, and Computing of Exploratory Data Analysis. Duxbury, 1981.
3.6
Exercises
1. Obtain one of the data sets available at the UCI Machine Learning Repository and apply as many of the diﬀerent visualization techniques described in the chapter as possible. The bibliographic notes and book Web site provide pointers to visualization software.
141
Chapter 3
Exploring Data
2. Identify at least two advantages and two disadvantages of using color to visually represent information. 3. What are the arrangement issues that arise with respect to threedimensional plots? 4. Discuss the advantages and disadvantages of using sampling to reduce the number of data objects that need to be displayed. Would simple random sampling (without replacement) be a good approach to sampling? Why or why not? 5. Describe how you would create visualizations to display information that describes the following types of systems. (a) Computer networks. Be sure to include both the static aspects of the network, such as connectivity, and the dynamic aspects, such as traﬃc. (b) The distribution of speciﬁc plant and animal species around the world for a speciﬁc moment in time. (c) The use of computer resources, such as processor time, main memory, and disk, for a set of benchmark database programs. (d) The change in occupation of workers in a particular country over the last thirty years. Assume that you have yearly information about each person that also includes gender and level of education. Be sure to address the following issues: • Representation. How will you map objects, attributes, and relationships to visual elements? • Arrangement. Are there any special considerations that need to be taken into account with respect to how visual elements are displayed? Speciﬁc examples might be the choice of viewpoint, the use of transparency, or the separation of certain groups of objects. • Selection. How will you handle a large number of attributes and data objects? 6. Describe one advantage and one disadvantage of a stem and leaf plot with respect to a standard histogram. 7. How might you address the problem that a histogram depends on the number and location of the bins? 8. Describe how a box plot can give information about whether the value of an attribute is symmetrically distributed. What can you say about the symmetry of the distributions of the attributes shown in Figure 3.11? 9. Compare sepal length, sepal width, petal length, and petal width, using Figure 3.12.
142
3.6
Exercises
10. Comment on the use of a box plot to explore a data set with four attributes: age, weight, height, and income. 11. Give a possible explanation as to why most of the values of petal length and width fall in the buckets along the diagonal in Figure 3.9. 12. Use Figures 3.14 and 3.15 to identify a characteristic shared by the petal width and petal length attributes. 13. Simple line plots, such as that displayed in Figure 2.12 on page 56, which shows two time series, can be used to eﬀectively display highdimensional data. For example, in Figure 2.12 it is easy to tell that the frequencies of the two time series are diﬀerent. What characteristic of time series allows the eﬀective visualization of highdimensional data? 14. Describe the types of situations that produce sparse or dense data cubes. Illustrate with examples other than those used in the book. 15. How might you extend the notion of multidimensional data analysis so that the target variable is a qualitative variable? In other words, what sorts of summary statistics or data visualizations would be of interest? 16. Construct a data cube from Table 3.14. Is this a dense or sparse data cube? If it is sparse, identify the cells that empty. Table 3.14. Fact table for Exercise 16.
Product ID 1 1 2 2
Location ID 1 3 1 2
Number Sold 10 6 5 22
17. Discuss the diﬀerences between dimensionality reduction based on aggregation and dimensionality reduction based on techniques such as PCA and SVD.
143
144
4 Classification: Basic Concepts, Decision Trees, and Model Evaluation Classiﬁcation, which is the task of assigning objects to one of several predeﬁned categories, is a pervasive problem that encompasses many diverse applications. Examples include detecting spam email messages based upon the message header and content, categorizing cells as malignant or benign based upon the results of MRI scans, and classifying galaxies based upon their shapes (see Figure 4.1).
(a) A spiral galaxy.
(b) An elliptical galaxy.
Figure 4.1. Classification of galaxies. The images are from the NASA website.
From Chapter 4 of Introduction to Data Mining, First Edition. PangNing Tan, Michael Steinbach, Vipin Kumar. Copyright © 2006 by Pearson Education, Inc. All rights reserved.
145
Chapter 4
Classiﬁcation
Input Attribute set (x)
Output Classification model
Class label (y)
Figure 4.2. Classification as the task of mapping an input attribute set x into its class label y.
This chapter introduces the basic concepts of classiﬁcation, describes some of the key issues such as model overﬁtting, and presents methods for evaluating and comparing the performance of a classiﬁcation technique. While it focuses mainly on a technique known as decision tree induction, most of the discussion in this chapter is also applicable to other classiﬁcation techniques, many of which are covered in Chapter 5.
4.1
Preliminaries
The input data for a classiﬁcation task is a collection of records. Each record, also known as an instance or example, is characterized by a tuple (x, y), where x is the attribute set and y is a special attribute, designated as the class label (also known as category or target attribute). Table 4.1 shows a sample data set used for classifying vertebrates into one of the following categories: mammal, bird, ﬁsh, reptile, or amphibian. The attribute set includes properties of a vertebrate such as its body temperature, skin cover, method of reproduction, ability to ﬂy, and ability to live in water. Although the attributes presented in Table 4.1 are mostly discrete, the attribute set can also contain continuous features. The class label, on the other hand, must be a discrete attribute. This is a key characteristic that distinguishes classiﬁcation from regression, a predictive modeling task in which y is a continuous attribute. Regression techniques are covered in Appendix D. Deﬁnition 4.1 (Classiﬁcation). Classiﬁcation is the task of learning a target function f that maps each attribute set x to one of the predeﬁned class labels y. The target function is also known informally as a classiﬁcation model. A classiﬁcation model is useful for the following purposes. Descriptive Modeling A classiﬁcation model can serve as an explanatory tool to distinguish between objects of diﬀerent classes. For example, it would be useful—for both biologists and others—to have a descriptive model that
146
4.1
Preliminaries
Table 4.1. The vertebrate data set. Name human python salmon whale frog komodo dragon bat pigeon cat leopard shark turtle penguin porcupine eel salamander
Body Temperature warmblooded coldblooded coldblooded warmblooded coldblooded coldblooded
Skin Cover hair scales scales hair none scales
Gives Birth yes no no yes no no
Aquatic Creature no no yes yes semi no
Aerial Creature no no no no no no
Has Legs yes no no no yes yes
Hibernates no yes no no yes no
Class Label mammal reptile ﬁsh mammal amphibian reptile
warmblooded warmblooded warmblooded coldblooded
hair feathers fur scales
yes no yes yes
no no no yes
yes yes no no
yes yes yes no
yes no no no
mammal bird mammal ﬁsh
coldblooded warmblooded warmblooded coldblooded coldblooded
scales feathers quills scales none
no no yes no no
semi semi no yes semi
no no no no no
yes yes yes no yes
no no yes no yes
reptile bird mammal ﬁsh amphibian
summarizes the data shown in Table 4.1 and explains what features deﬁne a vertebrate as a mammal, reptile, bird, ﬁsh, or amphibian. Predictive Modeling A classiﬁcation model can also be used to predict the class label of unknown records. As shown in Figure 4.2, a classiﬁcation model can be treated as a black box that automatically assigns a class label when presented with the attribute set of an unknown record. Suppose we are given the following characteristics of a creature known as a gila monster: Name gila monster
Body Temperature coldblooded
Skin Cover scales
Gives Birth no
Aquatic Creature no
Aerial Creature no
Has Legs yes
Hibernates yes
Class Label ?
We can use a classiﬁcation model built from the data set shown in Table 4.1 to determine the class to which the creature belongs. Classiﬁcation techniques are most suited for predicting or describing data sets with binary or nominal categories. They are less eﬀective for ordinal categories (e.g., to classify a person as a member of high, medium, or lowincome group) because they do not consider the implicit order among the categories. Other forms of relationships, such as the subclass–superclass relationships among categories (e.g., humans and apes are primates, which in
147
Chapter 4
Classiﬁcation
turn, is a subclass of mammals) are also ignored. The remainder of this chapter focuses only on binary or nominal class labels.
4.2
General Approach to Solving a Classiﬁcation Problem
A classiﬁcation technique (or classiﬁer) is a systematic approach to building classiﬁcation models from an input data set. Examples include decision tree classiﬁers, rulebased classiﬁers, neural networks, support vector machines, and na¨ıve Bayes classiﬁers. Each technique employs a learning algorithm to identify a model that best ﬁts the relationship between the attribute set and class label of the input data. The model generated by a learning algorithm should both ﬁt the input data well and correctly predict the class labels of records it has never seen before. Therefore, a key objective of the learning algorithm is to build models with good generalization capability; i.e., models that accurately predict the class labels of previously unknown records. Figure 4.3 shows a general approach for solving classiﬁcation problems. First, a training set consisting of records whose class labels are known must Training Set Tid Attrib1 Attrib2 1 Yes Large 2 No Medium 3 No Small 4 Yes Medium 5 No Large 6 No Medium 7 Yes Large 8 No Small 9 No Medium 10 No Small
Attrib3 125K 100K 70K 120K 95K 60K 220K 85K 75K 90K
Class No No No No Yes No No Yes No Yes
Learning Algorithm
Induction Learn Model
Model Apply Model
Test Set Tid Attrib1 Attrib2 11 No Small 12 Yes Medium 13 Yes Large 14 No Small 15 No Large
Attrib3 55K 80K 110K 95K 67K
Class ? ? ? ? ?
Deduction
Figure 4.3. General approach for building a classification model.
148
4.2
General Approach to Solving a Classiﬁcation Problem Table 4.2. Confusion matrix for a 2class problem.
Actual Class
Class = 1 Class = 0
Predicted Class Class = 1 Class = 0 f11 f10 f01 f00
be provided. The training set is used to build a classiﬁcation model, which is subsequently applied to the test set, which consists of records with unknown class labels. Evaluation of the performance of a classiﬁcation model is based on the counts of test records correctly and incorrectly predicted by the model. These counts are tabulated in a table known as a confusion matrix. Table 4.2 depicts the confusion matrix for a binary classiﬁcation problem. Each entry fij in this table denotes the number of records from class i predicted to be of class j. For instance, f01 is the number of records from class 0 incorrectly predicted as class 1. Based on the entries in the confusion matrix, the total number of correct predictions made by the model is (f11 + f00 ) and the total number of incorrect predictions is (f10 + f01 ). Although a confusion matrix provides the information needed to determine how well a classiﬁcation model performs, summarizing this information with a single number would make it more convenient to compare the performance of diﬀerent models. This can be done using a performance metric such as accuracy, which is deﬁned as follows: Accuracy =
f11 + f00 Number of correct predictions = . Total number of predictions f11 + f10 + f01 + f00
(4.1)
Equivalently, the performance of a model can be expressed in terms of its error rate, which is given by the following equation: Error rate =
Number of wrong predictions f10 + f01 = . Total number of predictions f11 + f10 + f01 + f00
(4.2)
Most classiﬁcation algorithms seek models that attain the highest accuracy, or equivalently, the lowest error rate when applied to the test set. We will revisit the topic of model evaluation in Section 4.5.
149
Chapter 4
4.3
Classiﬁcation
Decision Tree Induction
This section introduces a decision tree classiﬁer, which is a simple yet widely used classiﬁcation technique.
4.3.1
How a Decision Tree Works
To illustrate how classiﬁcation with a decision tree works, consider a simpler version of the vertebrate classiﬁcation problem described in the previous section. Instead of classifying the vertebrates into ﬁve distinct groups of species, we assign them to two categories: mammals and nonmammals. Suppose a new species is discovered by scientists. How can we tell whether it is a mammal or a nonmammal? One approach is to pose a series of questions about the characteristics of the species. The ﬁrst question we may ask is whether the species is cold or warmblooded. If it is coldblooded, then it is deﬁnitely not a mammal. Otherwise, it is either a bird or a mammal. In the latter case, we need to ask a followup question: Do the females of the species give birth to their young? Those that do give birth are deﬁnitely mammals, while those that do not are likely to be nonmammals (with the exception of egglaying mammals such as the platypus and spiny anteater). The previous example illustrates how we can solve a classiﬁcation problem by asking a series of carefully crafted questions about the attributes of the test record. Each time we receive an answer, a followup question is asked until we reach a conclusion about the class label of the record. The series of questions and their possible answers can be organized in the form of a decision tree, which is a hierarchical structure consisting of nodes and directed edges. Figure 4.4 shows the decision tree for the mammal classiﬁcation problem. The tree has three types of nodes: • A root node that has no incoming edges and zero or more outgoing edges. • Internal nodes, each of which has exactly one incoming edge and two or more outgoing edges. • Leaf or terminal nodes, each of which has exactly one incoming edge and no outgoing edges. In a decision tree, each leaf node is assigned a class label. The nonterminal nodes, which include the root and other internal nodes, contain attribute test conditions to separate records that have diﬀerent characteristics. For example, the root node shown in Figure 4.4 uses the attribute Body
150
4.3
Decision Tree Induction
Body Temperature Internal node
Warm
Gives Birth
Yes
Mammals
Root node Cold
Nonmammals
No
Nonmammals
Leaf nodes
Figure 4.4. A decision tree for the mammal classification problem.
Temperature to separate warmblooded from coldblooded vertebrates. Since all coldblooded vertebrates are nonmammals, a leaf node labeled Nonmammals is created as the right child of the root node. If the vertebrate is warmblooded, a subsequent attribute, Gives Birth, is used to distinguish mammals from other warmblooded creatures, which are mostly birds. Classifying a test record is straightforward once a decision tree has been constructed. Starting from the root node, we apply the test condition to the record and follow the appropriate branch based on the outcome of the test. This will lead us either to another internal node, for which a new test condition is applied, or to a leaf node. The class label associated with the leaf node is then assigned to the record. As an illustration, Figure 4.5 traces the path in the decision tree that is used to predict the class label of a ﬂamingo. The path terminates at a leaf node labeled Nonmammals.
4.3.2
How to Build a Decision Tree
In principle, there are exponentially many decision trees that can be constructed from a given set of attributes. While some of the trees are more accurate than others, ﬁnding the optimal tree is computationally infeasible because of the exponential size of the search space. Nevertheless, eﬃcient algorithms have been developed to induce a reasonably accurate, albeit suboptimal, decision tree in a reasonable amount of time. These algorithms usually employ a greedy strategy that grows a decision tree by making a series of locally op
151
Chapter 4
Classiﬁcation
Unlabeled Name data Flamingo
Body temperature Warm
Gives Birth No
Body Temperature Warm
Mammals
Class ?
Nonmammals Cold
Nonmammals
Gives Birth
Yes
... ...
No
Nonmammals
Figure 4.5. Classifying an unlabeled vertebrate. The dashed lines represent the outcomes of applying various attribute test conditions on the unlabeled vertebrate. The vertebrate is eventually assigned to the Nonmammal class.
timum decisions about which attribute to use for partitioning the data. One such algorithm is Hunt’s algorithm, which is the basis of many existing decision tree induction algorithms, including ID3, C4.5, and CART. This section presents a highlevel discussion of Hunt’s algorithm and illustrates some of its design issues. Hunt’s Algorithm In Hunt’s algorithm, a decision tree is grown in a recursive fashion by partitioning the training records into successively purer subsets. Let Dt be the set of training records that are associated with node t and y = {y1 , y2 , . . . , yc } be the class labels. The following is a recursive deﬁnition of Hunt’s algorithm. Step 1: If all the records in Dt belong to the same class yt , then t is a leaf node labeled as yt . Step 2: If Dt contains records that belong to more than one class, an attribute test condition is selected to partition the records into smaller subsets. A child node is created for each outcome of the test condition and the records in Dt are distributed to the children based on the outcomes. The algorithm is then recursively applied to each child node.
152
4.3
Decision Tree Induction
al
ry
na
bi
Tid 1 2 3 4 5 6 7 8 9 10
Home Owner Yes No No Yes No No Yes No No No
us
ic or
g
te
ca
Marital Status Single Married Single Married Divorced Married Divorced Single Married Single
uo tin
n
co
Annual Income 125K 100K 70K 120K 95K 60K 220K 85K 75K 90K
ss
a cl
Defaulted Borrower No No No No Yes No No Yes No Yes
Figure 4.6. Training set for predicting borrowers who will default on loan payments.
To illustrate how the algorithm works, consider the problem of predicting whether a loan applicant will repay her loan obligations or become delinquent, subsequently defaulting on her loan. A training set for this problem can be constructed by examining the records of previous borrowers. In the example shown in Figure 4.6, each record contains the personal information of a borrower along with a class label indicating whether the borrower has defaulted on loan payments. The initial tree for the classiﬁcation problem contains a single node with class label Defaulted = No (see Figure 4.7(a)), which means that most of the borrowers successfully repaid their loans. The tree, however, needs to be reﬁned since the root node contains records from both classes. The records are subsequently divided into smaller subsets based on the outcomes of the Home Owner test condition, as shown in Figure 4.7(b). The justiﬁcation for choosing this attribute test condition will be discussed later. For now, we will assume that this is the best criterion for splitting the data at this point. Hunt’s algorithm is then applied recursively to each child of the root node. From the training set given in Figure 4.6, notice that all borrowers who are home owners successfully repaid their loans. The left child of the root is therefore a leaf node labeled Defaulted = No (see Figure 4.7(b)). For the right child, we need to continue applying the recursive step of Hunt’s algorithm until all the records belong to the same class. The trees resulting from each recursive step are shown in Figures 4.7(c) and (d).
153
Chapter 4
Classiﬁcation Home Owner Yes
No
Defaulted = No Defaulted = No (a)
Defaulted = No (b) Home Owner
Yes Home Owner Yes Defaulted = No
No Marital Status
Defaulted = No Single, Divorced
No Marital Status
Single, Divorced
Annual Income Married
Defaulted = Yes (c)
Married
Defaulted = No
< 80K Defaulted = No
Defaulted = No >= 80K
Defaulted = Yes (d)
Figure 4.7. Hunt’s algorithm for inducing decision trees.
Hunt’s algorithm will work if every combination of attribute values is present in the training data and each combination has a unique class label. These assumptions are too stringent for use in most practical situations. Additional conditions are needed to handle the following cases: 1. It is possible for some of the child nodes created in Step 2 to be empty; i.e., there are no records associated with these nodes. This can happen if none of the training records have the combination of attribute values associated with such nodes. In this case the node is declared a leaf node with the same class label as the majority class of training records associated with its parent node. 2. In Step 2, if all the records associated with Dt have identical attribute values (except for the class label), then it is not possible to split these records any further. In this case, the node is declared a leaf node with the same class label as the majority class of training records associated with this node.
154
4.3
Decision Tree Induction
Design Issues of Decision Tree Induction A learning algorithm for inducing decision trees must address the following two issues. 1. How should the training records be split? Each recursive step of the treegrowing process must select an attribute test condition to divide the records into smaller subsets. To implement this step, the algorithm must provide a method for specifying the test condition for diﬀerent attribute types as well as an objective measure for evaluating the goodness of each test condition. 2. How should the splitting procedure stop? A stopping condition is needed to terminate the treegrowing process. A possible strategy is to continue expanding a node until either all the records belong to the same class or all the records have identical attribute values. Although both conditions are suﬃcient to stop any decision tree induction algorithm, other criteria can be imposed to allow the treegrowing procedure to terminate earlier. The advantages of early termination will be discussed later in Section 4.4.5.
4.3.3
Methods for Expressing Attribute Test Conditions
Decision tree induction algorithms must provide a method for expressing an attribute test condition and its corresponding outcomes for diﬀerent attribute types. Binary Attributes The test condition for a binary attribute generates two potential outcomes, as shown in Figure 4.8.
Body Temperature
Warmblooded
Coldblooded
Figure 4.8. Test condition for binary attributes.
155
Chapter 4
Classiﬁcation Marital Status
Single
Divorced
Married
(a) Multiway split
Marital Status
Marital Status OR
{Married}
{Single, Divorced}
Marital Status OR
{Single}
{Married, Divorced}
{Single, Married}
{Divorced}
(b) Binary split {by grouping attribute values}
Figure 4.9. Test conditions for nominal attributes.
Nominal Attributes Since a nominal attribute can have many values, its test condition can be expressed in two ways, as shown in Figure 4.9. For a multiway split (Figure 4.9(a)), the number of outcomes depends on the number of distinct values for the corresponding attribute. For example, if an attribute such as marital status has three distinct values—single, married, or divorced—its test condition will produce a threeway split. On the other hand, some decision tree algorithms, such as CART, produce only binary splits by considering all 2k−1 − 1 ways of creating a binary partition of k attribute values. Figure 4.9(b) illustrates three diﬀerent ways of grouping the attribute values for marital status into two subsets. Ordinal Attributes Ordinal attributes can also produce binary or multiway splits. Ordinal attribute values can be grouped as long as the grouping does not violate the order property of the attribute values. Figure 4.10 illustrates various ways of splitting training records based on the Shirt Size attribute. The groupings shown in Figures 4.10(a) and (b) preserve the order among the attribute values, whereas the grouping shown in Figure 4.10(c) violates this property because it combines the attribute values Small and Large into
156
4.3 Shirt Size
Shirt Size
{Small, Medium}
Decision Tree Induction
{Large, Extra Large} (a)
{Small}
Shirt Size
{Medium, Large, Extra Large}
{Small, Large}
(b)
{Medium, Extra Large} (c)
Figure 4.10. Different ways of grouping ordinal attribute values.
the same partition while Medium and Extra Large are combined into another partition. Continuous Attributes For continuous attributes, the test condition can be expressed as a comparison test (A < v) or (A ≥ v) with binary outcomes, or a range query with outcomes of the form vi ≤ A < vi+1 , for i = 1, . . . , k. The diﬀerence between these approaches is shown in Figure 4.11. For the binary case, the decision tree algorithm must consider all possible split positions v, and it selects the one that produces the best partition. For the multiway split, the algorithm must consider all possible ranges of continuous values. One approach is to apply the discretization strategies described in Section 2.3.6 on page 57. After discretization, a new ordinal value will be assigned to each discretized interval. Adjacent intervals can also be aggregated into wider ranges as long as the order property is preserved. Annual Income > 80K Yes
No
Annual Income < 10K
> 80K
{10K, 25K} {25K, 50K} {50K, 80K} (a)
(b)
Figure 4.11. Test condition for continuous attributes.
157
Chapter 4
Classiﬁcation Car Type
Gender Male
Female
Customer ID Luxury
Family
v1
Sports C0: 6 C1: 4
C0: 4 C1: 6 (a)
C0:1 C1: 3
C0: 8 C1: 0
v10
C0: 1 . . . C0: 1 C1: 0 C1: 0
C0: 1 C1: 7
(b)
v20
v11 C0: 0 C1: 1
. . . C0: 0 C1: 1
(c)
Figure 4.12. Multiway versus binary splits.
4.3.4
Measures for Selecting the Best Split
There are many measures that can be used to determine the best way to split the records. These measures are deﬁned in terms of the class distribution of the records before and after splitting. Let p(it) denote the fraction of records belonging to class i at a given node t. We sometimes omit the reference to node t and express the fraction as pi . In a twoclass problem, the class distribution at any node can be written as (p0 , p1 ), where p1 = 1 − p0 . To illustrate, consider the test conditions shown in Figure 4.12. The class distribution before splitting is (0.5, 0.5) because there are an equal number of records from each class. If we split the data using the Gender attribute, then the class distributions of the child nodes are (0.6, 0.4) and (0.4, 0.6), respectively. Although the classes are no longer evenly distributed, the child nodes still contain records from both classes. Splitting on the second attribute, Car Type, will result in purer partitions. The measures developed for selecting the best split are often based on the degree of impurity of the child nodes. The smaller the degree of impurity, the more skewed the class distribution. For example, a node with class distribution (0, 1) has zero impurity, whereas a node with uniform class distribution (0.5, 0.5) has the highest impurity. Examples of impurity measures include Entropy(t) = −
c−1
p(it) log2 p(it),
i=0 c−1
Gini(t) = 1 −
(4.3)
[p(it)]2 ,
(4.4)
Classiﬁcation error(t) = 1 − max[p(it)],
(4.5)
i=0 i
where c is the number of classes and 0 log2 0 = 0 in entropy calculations.
158
4.3
Decision Tree Induction
1 0.9
Entropy
0.8 0.7 0.6 0.5 Gini 0.4 0.3 Misclassification error
0.2 0.1 0 0
0.1
0.2
0.3
0.4
0.5 p
0.6
0.7
0.8
0.9
1
Figure 4.13. Comparison among the impurity measures for binary classification problems.
Figure 4.13 compares the values of the impurity measures for binary classiﬁcation problems. p refers to the fraction of records that belong to one of the two classes. Observe that all three measures attain their maximum value when the class distribution is uniform (i.e., when p = 0.5). The minimum values for the measures are attained when all the records belong to the same class (i.e., when p equals 0 or 1). We next provide several examples of computing the diﬀerent impurity measures. Node N1 Class=0 Class=1
Count 0 6
Gini = 1 − (0/6)2 − (6/6)2 = 0 Entropy = −(0/6) log2 (0/6) − (6/6) log2 (6/6) = 0 Error = 1 − max[0/6, 6/6] = 0
Node N2 Class=0 Class=1
Count 1 5
Gini = 1 − (1/6)2 − (5/6)2 = 0.278 Entropy = −(1/6) log2 (1/6) − (5/6) log2 (5/6) = 0.650 Error = 1 − max[1/6, 5/6] = 0.167
Node N3 Class=0 Class=1
Count 3 3
Gini = 1 − (3/6)2 − (3/6)2 = 0.5 Entropy = −(3/6) log2 (3/6) − (3/6) log2 (3/6) = 1 Error = 1 − max[3/6, 3/6] = 0.5
159
Chapter 4
Classiﬁcation
The preceding examples, along with Figure 4.13, illustrate the consistency among diﬀerent impurity measures. Based on these calculations, node N1 has the lowest impurity value, followed by N2 and N3 . Despite their consistency, the attribute chosen as the test condition may vary depending on the choice of impurity measure, as will be shown in Exercise 3 on page 198. To determine how well a test condition performs, we need to compare the degree of impurity of the parent node (before splitting) with the degree of impurity of the child nodes (after splitting). The larger their diﬀerence, the better the test condition. The gain, ∆, is a criterion that can be used to determine the goodness of a split: ∆ = I(parent) −
k N (vj ) j=1
N
I(vj ),
(4.6)
where I(·) is the impurity measure of a given node, N is the total number of records at the parent node, k is the number of attribute values, and N (vj ) is the number of records associated with the child node, vj . Decision tree induction algorithms often choose a test condition that maximizes the gain ∆. Since I(parent) is the same for all test conditions, maximizing the gain is equivalent to minimizing the weighted average impurity measures of the child nodes. Finally, when entropy is used as the impurity measure in Equation 4.6, the diﬀerence in entropy is known as the information gain, ∆info . Splitting of Binary Attributes Consider the diagram shown in Figure 4.14. Suppose there are two ways to split the data into smaller subsets. Before splitting, the Gini index is 0.5 since there are an equal number of records from both classes. If attribute A is chosen to split the data, the Gini index for node N1 is 0.4898, and for node N2, it is 0.480. The weighted average of the Gini index for the descendent nodes is (7/12) × 0.4898 + (5/12) × 0.480 = 0.486. Similarly, we can show that the weighted average of the Gini index for attribute B is 0.375. Since the subsets for attribute B have a smaller Gini index, it is preferred over attribute A. Splitting of Nominal Attributes As previously noted, a nominal attribute can produce either binary or multiway splits, as shown in Figure 4.15. The computation of the Gini index for a binary split is similar to that shown for determining binary attributes. For the ﬁrst binary grouping of the Car Type attribute, the Gini index of {Sports,
160
4.3
Decision Tree Induction
Parent C0
6
C1
6
Gini = 0.500 A
B
Yes
No
Yes
Node N2
Node N1
No
Node N1
N1 N2
Node N2
N1 N2
C0
4
2
C0
1
5
C1
3
3
C1
4
2
Gini = 0.486
Gini = 0.375
Figure 4.14. Splitting binary attributes.
Car Type {Sports, Luxury}
C0 C1 Gini
{Family}
Car Type
Car Type
{Family, Luxury}
Family
{Sports}
Luxury
Sports
Car Type
Car Type
Car Type
{Sports, {Family} Luxury}
{Family, {Sports} Luxury}
Family Sports Luxury
9 7
1 3 0.468
C0 C1 Gini
8 0
2 10 0.167
C0 C1
1 3
Gini
(a) Binary split
8 0
1 7
0.163 (b) Multiway split
Figure 4.15. Splitting nominal attributes.
Luxury} is 0.4922 and the Gini index of {Family} is 0.3750. The weighted average Gini index for the grouping is equal to 16/20 × 0.4922 + 4/20 × 0.3750 = 0.468. Similarly, for the second binary grouping of {Sports} and {Family, Luxury}, the weighted average Gini index is 0.167. The second grouping has a lower Gini index because its corresponding subsets are much purer.
161
Chapter 4 Class Sorted Values Split Positions
No 60
Classiﬁcation No 70
No
Yes
No
No
No
No
75
Annual Income 85 90 95 100
Yes
Yes
120
125
220
55 65 72 80 87 92 97 110 122 172 230
Yes
0 3
0 3
0 3
0 3
1 2
2 1
3 0
3 0
3
0
3
0
3
0
No
0 7
1 6
2 5
3 4
3 4
3 4
3 4
4 3
5
2
6
1
7
0
Gini
0.420 0.400 0.375 0.343 0.417 0.400 0.300 0.343 0.375
0.400
0.420
Figure 4.16. Splitting continuous attributes.
For the multiway split, the Gini index is computed for every attribute value. Since Gini({Family}) = 0.375, Gini({Sports}) = 0, and Gini({Luxury}) = 0.219, the overall Gini index for the multiway split is equal to 4/20 × 0.375 + 8/20 × 0 + 8/20 × 0.219 = 0.163. The multiway split has a smaller Gini index compared to both twoway splits. This result is not surprising because the twoway split actually merges some of the outcomes of a multiway split, and thus, results in less pure subsets. Splitting of Continuous Attributes Consider the example shown in Figure 4.16, in which the test condition Annual Income ≤ v is used to split the training records for the loan default classiﬁcation problem. A bruteforce method for ﬁnding v is to consider every value of the attribute in the N records as a candidate split position. For each candidate v, the data set is scanned once to count the number of records with annual income less than or greater than v. We then compute the Gini index for each candidate and choose the one that gives the lowest value. This approach is computationally expensive because it requires O(N ) operations to compute the Gini index at each candidate split position. Since there are N candidates, the overall complexity of this task is O(N 2 ). To reduce the complexity, the training records are sorted based on their annual income, a computation that requires O(N log N ) time. Candidate split positions are identiﬁed by taking the midpoints between two adjacent sorted values: 55, 65, 72, and so on. However, unlike the bruteforce approach, we do not have to examine all N records when evaluating the Gini index of a candidate split position. For the ﬁrst candidate, v = 55, none of the records has annual income less than $55K. As a result, the Gini index for the descendent node with Annual
162
4.3
Decision Tree Induction
Income < $55K is zero. On the other hand, the number of records with annual income greater than or equal to $55K is 3 (for class Yes) and 7 (for class No), respectively. Thus, the Gini index for this node is 0.420. The overall Gini index for this candidate split position is equal to 0 × 0 + 1 × 0.420 = 0.420. For the second candidate, v = 65, we can determine its class distribution by updating the distribution of the previous candidate. More speciﬁcally, the new distribution is obtained by examining the class label of the record with the lowest annual income (i.e., $60K). Since the class label for this record is No, the count for class No is increased from 0 to 1 (for Annual Income ≤ $65K) and is decreased from 7 to 6 (for Annual Income > $65K). The distribution for class Yes remains unchanged. The new weightedaverage Gini index for this candidate split position is 0.400. This procedure is repeated until the Gini index values for all candidates are computed, as shown in Figure 4.16. The best split position corresponds to the one that produces the smallest Gini index, i.e., v = 97. This procedure is less expensive because it requires a constant amount of time to update the class distribution at each candidate split position. It can be further optimized by considering only candidate split positions located between two adjacent records with diﬀerent class labels. For example, because the ﬁrst three sorted records (with annual incomes $60K, $70K, and $75K) have identical class labels, the best split position should not reside between $60K and $75K. Therefore, the candidate split positions at v = $55K, $65K, $72K, $87K, $92K, $110K, $122K, $172K, and $230K are ignored because they are located between two adjacent records with the same class labels. This approach allows us to reduce the number of candidate split positions from 11 to 2. Gain Ratio Impurity measures such as entropy and Gini index tend to favor attributes that have a large number of distinct values. Figure 4.12 shows three alternative test conditions for partitioning the data set given in Exercise 2 on page 198. Comparing the ﬁrst test condition, Gender, with the second, Car Type, it is easy to see that Car Type seems to provide a better way of splitting the data since it produces purer descendent nodes. However, if we compare both conditions with Customer ID, the latter appears to produce purer partitions. Yet Customer ID is not a predictive attribute because its value is unique for each record. Even in a less extreme situation, a test condition that results in a large number of outcomes may not be desirable because the number of records associated with each partition is too small to enable us to make any reliable predictions.
163
Chapter 4
Classiﬁcation
There are two strategies for overcoming this problem. The ﬁrst strategy is to restrict the test conditions to binary splits only. This strategy is employed by decision tree algorithms such as CART. Another strategy is to modify the splitting criterion to take into account the number of outcomes produced by the attribute test condition. For example, in the C4.5 decision tree algorithm, a splitting criterion known as gain ratio is used to determine the goodness of a split. This criterion is deﬁned as follows: Gain ratio =
∆info . Split Info
(4.7)
Here, Split Info = − ki=1 P (vi ) log2 P (vi ) and k is the total number of splits. For example, if each attribute value has the same number of records, then ∀i : P (vi ) = 1/k and the split information would be equal to log2 k. This example suggests that if an attribute produces a large number of splits, its split information will also be large, which in turn reduces its gain ratio.
4.3.5
Algorithm for Decision Tree Induction
A skeleton decision tree induction algorithm called TreeGrowth is shown in Algorithm 4.1. The input to this algorithm consists of the training records E and the attribute set F . The algorithm works by recursively selecting the best attribute to split the data (Step 7) and expanding the leaf nodes of the Algorithm 4.1 A skeleton decision tree induction algorithm. TreeGrowth (E, F ) 1: if stopping cond(E,F ) = true then 2: leaf = createNode(). 3: leaf.label = Classify(E). 4: return leaf . 5: else 6: root = createNode(). 7: root.test cond = find best split(E, F ). 8: let V = {vv is a possible outcome of root.test cond }. 9: for each v ∈ V do 10: Ev = {e  root.test cond(e) = v and e ∈ E}. 11: child = TreeGrowth(Ev , F ). 12: add child as descendent of root and label the edge (root → child) as v. 13: end for 14: end if 15: return root.
164
4.3
Decision Tree Induction
tree (Steps 11 and 12) until the stopping criterion is met (Step 1). The details of this algorithm are explained below: 1. The createNode() function extends the decision tree by creating a new node. A node in the decision tree has either a test condition, denoted as node.test cond, or a class label, denoted as node.label. 2. The find best split() function determines which attribute should be selected as the test condition for splitting the training records. As previously noted, the choice of test condition depends on which impurity measure is used to determine the goodness of a split. Some widely used measures include entropy, the Gini index, and the χ2 statistic. 3. The Classify() function determines the class label to be assigned to a leaf node. For each leaf node t, let p(it) denote the fraction of training records from class i associated with the node t. In most cases, the leaf node is assigned to the class that has the majority number of training records: (4.8) leaf.label = argmax p(it), i
where the argmax operator returns the argument i that maximizes the expression p(it). Besides providing the information needed to determine the class label of a leaf node, the fraction p(it) can also be used to estimate the probability that a record assigned to the leaf node t belongs to class i. Sections 5.7.2 and 5.7.3 describe how such probability estimates can be used to determine the performance of a decision tree under diﬀerent cost functions. 4. The stopping cond() function is used to terminate the treegrowing process by testing whether all the records have either the same class label or the same attribute values. Another way to terminate the recursive function is to test whether the number of records have fallen below some minimum threshold. After building the decision tree, a treepruning step can be performed to reduce the size of the decision tree. Decision trees that are too large are susceptible to a phenomenon known as overﬁtting. Pruning helps by trimming the branches of the initial tree in a way that improves the generalization capability of the decision tree. The issues of overﬁtting and tree pruning are discussed in more detail in Section 4.4.
165
Chapter 4 Session IP Address
Classiﬁcation
Timestamp Request Requested Web Page Method
Protocol Status Number of Bytes
1
160.11.11.11 08/Aug/2004 10:15:21
GET
http://www.cs.umn.edu/ HTTP/1.1 ~kumar
200
1
160.11.11.11 08/Aug/2004 10:15:34
GET
http://www.cs.umn.edu/ HTTP/1.1 ~kumar/MINDS
200
1
160.11.11.11 08/Aug/2004 10:15:41
GET
200
1
160.11.11.11 08/Aug/2004 10:16:11
GET
http://www.cs.umn.edu/ HTTP/1.1 ~kumar/MINDS/MINDS _papers.htm http://www.cs.umn.edu/ HTTP/1.1 ~kumar/papers/papers. html http://www.cs.umn.edu/ HTTP/1.0 ~steinbac
2
35.9.2.2
08/Aug/2004 10:16:15
GET
200 200
Referrer
User Agent
Mozilla/4.0 (compatible; MSIE 6.0; Windows NT 5.0) 41378 http://www.cs.umn.edu/ Mozilla/4.0 (compatible; MSIE 6.0; ~kumar Windows NT 5.0) 1018516 http://www.cs.umn.edu/ Mozilla/4.0 (compatible; MSIE 6.0; ~kumar/MINDS Windows NT 5.0) 7463 http://www.cs.umn.edu/ Mozilla/4.0 (compatible; MSIE 6.0; ~kumar Windows NT 5.0) 3149 Mozilla/5.0 (Windows; U; Windows NT 5.1; enUS; rv:1.7) Gecko/20040616 6424
(a) Example of a Web server log.
http://www.cs.umn.edu/~kumar
MINDS papers/papers.html
MINDS/MINDS_papers.htm (b) Graph of a Web session.
Attribute Name totalPages ImagePages TotalTime RepeatedAccess ErrorRequest GET POST HEAD Breadth Depth MultilP MultiAgent
Description Total number of pages retrieved in a Web session Total number of image pages retrieved in a Web session Total amount of time spent by Web site visitor The same page requested more than once in a Web session Errors in requesting for Web pages Percentage of requests made using GET method Percentage of requests made using POST method Percentage of requests made using HEAD method Breadth of Web traversal Depth of Web traversal Session with multiple IP addresses Session with multiple user agents
(c) Derived attributes for Web robot detection.
Figure 4.17. Input data for Web robot detection.
4.3.6
An Example: Web Robot Detection
Web usage mining is the task of applying data mining techniques to extract useful patterns from Web access logs. These patterns can reveal interesting characteristics of site visitors; e.g., people who repeatedly visit a Web site and view the same product description page are more likely to buy the product if certain incentives such as rebates or free shipping are oﬀered. In Web usage mining, it is important to distinguish accesses made by human users from those due to Web robots. A Web robot (also known as a Web crawler) is a software program that automatically locates and retrieves information from the Internet by following the hyperlinks embedded in Web pages. These programs are deployed by search engine portals to gather the documents necessary for indexing the Web. Web robot accesses must be discarded before applying Web mining techniques to analyze human browsing behavior.
166
4.3
Decision Tree Induction
This section describes how a decision tree classiﬁer can be used to distinguish between accesses by human users and those by Web robots. The input data was obtained from a Web server log, a sample of which is shown in Figure 4.17(a). Each line corresponds to a single page request made by a Web client (a user or a Web robot). The ﬁelds recorded in the Web log include the IP address of the client, timestamp of the request, Web address of the requested document, size of the document, and the client’s identity (via the user agent ﬁeld). A Web session is a sequence of requests made by a client during a single visit to a Web site. Each Web session can be modeled as a directed graph, in which the nodes correspond to Web pages and the edges correspond to hyperlinks connecting one Web page to another. Figure 4.17(b) shows a graphical representation of the ﬁrst Web session given in the Web server log. To classify the Web sessions, features are constructed to describe the characteristics of each session. Figure 4.17(c) shows some of the features used for the Web robot detection task. Among the notable features include the depth and breadth of the traversal. Depth determines the maximum distance of a requested page, where distance is measured in terms of the number of hyperlinks away from the entry point of the Web site. For example, the home page http://www.cs.umn.edu/∼kumar is assumed to be at depth 0, whereas http://www.cs.umn.edu/kumar/MINDS/MINDS papers.htm is located at depth 2. Based on the Web graph shown in Figure 4.17(b), the depth attribute for the ﬁrst session is equal to two. The breadth attribute measures the width of the corresponding Web graph. For example, the breadth of the Web session shown in Figure 4.17(b) is equal to two. The data set for classiﬁcation contains 2916 records, with equal numbers of sessions due to Web robots (class 1) and human users (class 0). 10% of the data were reserved for training while the remaining 90% were used for testing. The induced decision tree model is shown in Figure 4.18. The tree has an error rate equal to 3.8% on the training set and 5.3% on the test set. The model suggests that Web robots can be distinguished from human users in the following way: 1. Accesses by Web robots tend to be broad but shallow, whereas accesses by human users tend to be more focused (narrow but deep). 2. Unlike human users, Web robots seldom retrieve the image pages associated with a Web document. 3. Sessions due to Web robots tend to be long and contain a large number of requested pages.
167
Chapter 4
Classiﬁcation Decision Tree: depth = 1:  breadth> 7 : class 1  breadth 1:  MultiAgent = 0:   depth > 2: class 0   depth < 2:    MultilP = 1: class 0    MultilP = 0:     breadth 6:      RepeatedAccess 0.322: class 1  MultiAgent = 1:   totalPages 81: class 1
Figure 4.18. Decision tree model for Web robot detection.
4. Web robots are more likely to make repeated requests for the same document since the Web pages retrieved by human users are often cached by the browser.
4.3.7
Characteristics of Decision Tree Induction
The following is a summary of the important characteristics of decision tree induction algorithms. 1. Decision tree induction is a nonparametric approach for building classiﬁcation models. In other words, it does not require any prior assumptions regarding the type of probability distributions satisﬁed by the class and other attributes (unlike some of the techniques described in Chapter 5).
168
4.3
Decision Tree Induction
2. Finding an optimal decision tree is an NPcomplete problem. Many decision tree algorithms employ a heuristicbased approach to guide their search in the vast hypothesis space. For example, the algorithm presented in Section 4.3.5 uses a greedy, topdown, recursive partitioning strategy for growing a decision tree. 3. Techniques developed for constructing decision trees are computationally inexpensive, making it possible to quickly construct models even when the training set size is very large. Furthermore, once a decision tree has been built, classifying a test record is extremely fast, with a worstcase complexity of O(w), where w is the maximum depth of the tree. 4. Decision trees, especially smallersized trees, are relatively easy to interpret. The accuracies of the trees are also comparable to other classiﬁcation techniques for many simple data sets. 5. Decision trees provide an expressive representation for learning discretevalued functions. However, they do not generalize well to certain types of Boolean problems. One notable example is the parity function, whose value is 0 (1) when there is an odd (even) number of Boolean attributes with the value T rue. Accurate modeling of such a function requires a full decision tree with 2d nodes, where d is the number of Boolean attributes (see Exercise 1 on page 198). 6. Decision tree algorithms are quite robust to the presence of noise, especially when methods for avoiding overﬁtting, as described in Section 4.4, are employed. 7. The presence of redundant attributes does not adversely aﬀect the accuracy of decision trees. An attribute is redundant if it is strongly correlated with another attribute in the data. One of the two redundant attributes will not be used for splitting once the other attribute has been chosen. However, if the data set contains many irrelevant attributes, i.e., attributes that are not useful for the classiﬁcation task, then some of the irrelevant attributes may be accidently chosen during the treegrowing process, which results in a decision tree that is larger than necessary. Feature selection techniques can help to improve the accuracy of decision trees by eliminating the irrelevant attributes during preprocessing. We will investigate the issue of too many irrelevant attributes in Section 4.4.3.
169
Chapter 4
Classiﬁcation
8. Since most decision tree algorithms employ a topdown, recursive partitioning approach, the number of records becomes smaller as we traverse down the tree. At the leaf nodes, the number of records may be too small to make a statistically signiﬁcant decision about the class representation of the nodes. This is known as the data fragmentation problem. One possible solution is to disallow further splitting when the number of records falls below a certain threshold. 9. A subtree can be replicated multiple times in a decision tree, as illustrated in Figure 4.19. This makes the decision tree more complex than necessary and perhaps more diﬃcult to interpret. Such a situation can arise from decision tree implementations that rely on a single attribute test condition at each internal node. Since most of the decision tree algorithms use a divideandconquer partitioning strategy, the same test condition can be applied to diﬀerent parts of the attribute space, thus leading to the subtree replication problem. P
Q
S
0
R
0
Q
1
S
0
1
0
1
Figure 4.19. Tree replication problem. The same subtree can appear at different branches.
10. The test conditions described so far in this chapter involve using only a single attribute at a time. As a consequence, the treegrowing procedure can be viewed as the process of partitioning the attribute space into disjoint regions until each region contains records of the same class (see Figure 4.20). The border between two neighboring regions of diﬀerent classes is known as a decision boundary. Since the test condition involves only a single attribute, the decision boundaries are rectilinear; i.e., parallel to the “coordinate axes.” This limits the expressiveness of the
170
4.3
Decision Tree Induction
1 0.9 x < 0.43
0.8 0.7
Yes
No
y
0.6 0.5
y < 0.47
y < 0.33
0.4 Yes
0.3 0.2
0
0.1
0.2
0.3 0.4
0.5 x
0.6 0.7
0.8 0.9
Yes
:0 :4
:4 :0
0.1 0
No
:0 :3
No :4 :0
1
Figure 4.20. Example of a decision tree and its decision boundaries for a twodimensional data set.
1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Figure 4.21. Example of data set that cannot be partitioned optimally using test conditions involving single attributes.
decision tree representation for modeling complex relationships among continuous attributes. Figure 4.21 illustrates a data set that cannot be classiﬁed eﬀectively by a decision tree algorithm that uses test conditions involving only a single attribute at a time.
171
Chapter 4
Classiﬁcation
An oblique decision tree can be used to overcome this limitation because it allows test conditions that involve more than one attribute. The data set given in Figure 4.21 can be easily represented by an oblique decision tree containing a single node with test condition x + y < 1. Although such techniques are more expressive and can produce more compact trees, ﬁnding the optimal test condition for a given node can be computationally expensive. Constructive induction provides another way to partition the data into homogeneous, nonrectangular regions (see Section 2.3.5 on page 57). This approach creates composite attributes representing an arithmetic or logical combination of the existing attributes. The new attributes provide a better discrimination of the classes and are augmented to the data set prior to decision tree induction. Unlike the oblique decision tree approach, constructive induction is less expensive because it identiﬁes all the relevant combinations of attributes once, prior to constructing the decision tree. In contrast, an oblique decision tree must determine the right attribute combination dynamically, every time an internal node is expanded. However, constructive induction can introduce attribute redundancy in the data since the new attribute is a combination of several existing attributes. 11. Studies have shown that the choice of impurity measure has little eﬀect on the performance of decision tree induction algorithms. This is because many impurity measures are quite consistent with each other, as shown in Figure 4.13 on page 159. Indeed, the strategy used to prune the tree has a greater impact on the ﬁnal tree than the choice of impurity measure.
4.4
Model Overﬁtting
The errors committed by a classiﬁcation model are generally divided into two types: training errors and generalization errors. Training error, also known as resubstitution error or apparent error, is the number of misclassiﬁcation errors committed on training records, whereas generalization error is the expected error of the model on previously unseen records. Recall from Section 4.2 that a good classiﬁcation model must not only ﬁt the training data well, it must also accurately classify records it has never
172
4.4
Model Overﬁtting
Training set
20 18 16 14
x2
12 10 8 6 4 2 0
0
2
4
6
8
10 x1
12
14
16
18
20
Figure 4.22. Example of a data set with binary classes.
seen before. In other words, a good model must have low training error as well as low generalization error. This is important because a model that ﬁts the training data too well can have a poorer generalization error than a model with a higher training error. Such a situation is known as model overﬁtting. Overﬁtting Example in TwoDimensional Data For a more concrete example of the overﬁtting problem, consider the twodimensional data set shown in Figure 4.22. The data set contains data points that belong to two diﬀerent classes, denoted as class o and class +, respectively. The data points for the o class are generated from a mixture of three Gaussian distributions, while a uniform distribution is used to generate the data points for the + class. There are altogether 1200 points belonging to the o class and 1800 points belonging to the + class. 30% of the points are chosen for training, while the remaining 70% are used for testing. A decision tree classiﬁer that uses the Gini index as its impurity measure is then applied to the training set. To investigate the eﬀect of overﬁtting, diﬀerent levels of pruning are applied to the initial, fullygrown tree. Figure 4.23(b) shows the training and test error rates of the decision tree.
173
Chapter 4
Classiﬁcation
0.4 0.35
Training Error Test Error
Error Rate
0.3 0.25 0.2 0.15 0.1 0.05 0
0
50
100
150 200 Number of Nodes
250
300
Figure 4.23. Training and test error rates.
Notice that the training and test error rates of the model are large when the size of the tree is very small. This situation is known as model underﬁtting. Underﬁtting occurs because the model has yet to learn the true structure of the data. As a result, it performs poorly on both the training and the test sets. As the number of nodes in the decision tree increases, the tree will have fewer training and test errors. However, once the tree becomes too large, its test error rate begins to increase even though its training error rate continues to decrease. This phenomenon is known as model overﬁtting. To understand the overﬁtting phenomenon, note that the training error of a model can be reduced by increasing the model complexity. For example, the leaf nodes of the tree can be expanded until it perfectly ﬁts the training data. Although the training error for such a complex tree is zero, the test error can be large because the tree may contain nodes that accidently ﬁt some of the noise points in the training data. Such nodes can degrade the performance of the tree because they do not generalize well to the test examples. Figure 4.24 shows the structure of two decision trees with diﬀerent number of nodes. The tree that contains the smaller number of nodes has a higher training error rate, but a lower test error rate compared to the more complex tree. Overﬁtting and underﬁtting are two pathologies that are related to the model complexity. The remainder of this section examines some of the potential causes of model overﬁtting.
174
4.4
Model Overﬁtting x2 < 12.63 x2 < 17.35
x1 < 13.29 x1 < 2.15
x1 < 6.56
x2 < 12.63
x1 < 3.03
x2 < 1.38
x2 < 17.35
x1 < 13.29
x1 < 12.11
x1 < 6.78 x2 < 12.68
x1 < 6.99
x1 < 18.88
x1 < 2.72
x2 < 4.06
x1 < 2.15
x1 < 6.56
x2 < 17.14
x2 < 15.77
x2 < 12.89
x1 < 7.24
x2 < 8.64
x2 < 19.93 x1 < 7.24
x2 < 8.64
x2 < 13.80
x1 < 12.11
x2 < 1.38
x2 < 16.75
x1 < 18.88 x2 < 16.33
(a) Decision tree with 11 leaf nodes.
(b) Decision tree with 24 leaf nodes.
Figure 4.24. Decision trees with different model complexities.
4.4.1
Overﬁtting Due to Presence of Noise
Consider the training and test sets shown in Tables 4.3 and 4.4 for the mammal classiﬁcation problem. Two of the ten training records are mislabeled: bats and whales are classiﬁed as nonmammals instead of mammals. A decision tree that perfectly ﬁts the training data is shown in Figure 4.25(a). Although the training error for the tree is zero, its error rate on Table 4.3. An example training set for classifying mammals. Class labels with asterisk symbols represent mislabeled records. Name porcupine cat bat whale salamander komodo dragon python salmon eagle guppy
Body Temperature warmblooded warmblooded warmblooded warmblooded coldblooded coldblooded coldblooded coldblooded warmblooded coldblooded
Gives Birth yes yes yes yes no no no no no yes
Fourlegged yes yes no no yes yes no no no no
Hibernates yes no yes no yes no yes no no no
Class Label yes yes no∗ no∗ no no no no no no
175
Chapter 4
Classiﬁcation
Table 4.4. An example test set for classifying mammals. Name
Body Temperature warmblooded warmblooded warmblooded coldblooded coldblooded coldblooded coldblooded warmblooded warmblooded coldblooded
human pigeon elephant leopard shark turtle penguin eel dolphin spiny anteater gila monster
Gives Birth yes no yes yes no no no yes no no
Fourlegged no no yes no yes no no no yes yes
Body Temperature Warmblooded
Coldblooded
Gives Birth Yes
No Nonmammals
Fourlegged Yes
Mammals
Nonmammals
Hibernates no no no no no no no no yes yes
Class Label yes no yes no no no no yes yes no
Body Temperature Warmblooded
Coldblooded
Gives Birth Yes
Mammals
Nonmammals
No Nonmammals
No Nonmammals (a) Model M1
(b) Model M2
Figure 4.25. Decision tree induced from the data set shown in Table 4.3.
the test set is 30%. Both humans and dolphins were misclassiﬁed as nonmammals because their attribute values for Body Temperature, Gives Birth, and Fourlegged are identical to the mislabeled records in the training set. Spiny anteaters, on the other hand, represent an exceptional case in which the class label of a test record contradicts the class labels of other similar records in the training set. Errors due to exceptional cases are often unavoidable and establish the minimum error rate achievable by any classiﬁer.
176
4.4
Model Overﬁtting
In contrast, the decision tree M 2 shown in Figure 4.25(b) has a lower test error rate (10%) even though its training error rate is somewhat higher (20%). It is evident that the ﬁrst decision tree, M 1, has overﬁtted the training data because there is a simpler model with lower error rate on the test set. The Fourlegged attribute test condition in model M 1 is spurious because it ﬁts the mislabeled training records, which leads to the misclassiﬁcation of records in the test set.
4.4.2
Overﬁtting Due to Lack of Representative Samples
Models that make their classiﬁcation decisions based on a small number of training records are also susceptible to overﬁtting. Such models can be generated because of lack of representative samples in the training data and learning algorithms that continue to reﬁne their models even when few training records are available. We illustrate these eﬀects in the example below. Consider the ﬁve training records shown in Table 4.5. All of these training records are labeled correctly and the corresponding decision tree is depicted in Figure 4.26. Although its training error is zero, its error rate on the test set is 30%. Table 4.5. An example training set for classifying mammals. Name salamander guppy eagle poorwill platypus
Body Temperature coldblooded coldblooded warmblooded warmblooded warmblooded
Gives Birth no yes no no no
Fourlegged yes no no no yes
Hibernates yes no no yes yes
Class Label no no no no yes
Humans, elephants, and dolphins are misclassiﬁed because the decision tree classiﬁes all warmblooded vertebrates that do not hibernate as nonmammals. The tree arrives at this classiﬁcation decision because there is only one training record, which is an eagle, with such characteristics. This example clearly demonstrates the danger of making wrong predictions when there are not enough representative examples at the leaf nodes of a decision tree.
177
Chapter 4
Classiﬁcation Body Temperature Warmblooded
Hibernates Yes
Mammals
Nonmammals
No Nonmammals
Fourlegged Yes
Coldblooded
No Nonmammals
Figure 4.26. Decision tree induced from the data set shown in Table 4.5.
4.4.3
Overﬁtting and the Multiple Comparison Procedure
Model overﬁtting may arise in learning algorithms that employ a methodology known as multiple comparison procedure. To understand multiple comparison procedure, consider the task of predicting whether the stock market will rise or fall in the next ten trading days. If a stock analyst simply makes random guesses, the probability that her prediction is correct on any trading day is 0.5. However, the probability that she will predict correctly at least eight out of the ten times is 10 10 10 8 + 9 + 10 = 0.0547, 210 which seems quite unlikely. Suppose we are interested in choosing an investment advisor from a pool of ﬁfty stock analysts. Our strategy is to select the analyst who makes the most correct predictions in the next ten trading days. The ﬂaw in this strategy is that even if all the analysts had made their predictions in a random fashion, the probability that at least one of them makes at least eight correct predictions is 1 − (1 − 0.0547)50 = 0.9399, which is very high. Although each analyst has a low probability of predicting at least eight times correctly, putting them together, we have a high probability of ﬁnding an analyst who can do so. Furthermore, there is no guarantee in the
178
4.4
Model Overﬁtting
future that such an analyst will continue to make accurate predictions through random guessing. How does the multiple comparison procedure relate to model overﬁtting? Many learning algorithms explore a set of independent alternatives, {γi }, and then choose an alternative, γmax , that maximizes a given criterion function. The algorithm will add γmax to the current model in order to improve its overall performance. This procedure is repeated until no further improvement is observed. As an example, during decision tree growing, multiple tests are performed to determine which attribute can best split the training data. The attribute that leads to the best split is chosen to extend the tree as long as the observed improvement is statistically signiﬁcant. Let T0 be the initial decision tree and Tx be the new tree after inserting an internal node for attribute x. In principle, x can be added to the tree if the observed gain, ∆(T0 , Tx ), is greater than some predeﬁned threshold α. If there is only one attribute test condition to be evaluated, then we can avoid inserting spurious nodes by choosing a large enough value of α. However, in practice, more than one test condition is available and the decision tree algorithm must choose the best attribute xmax from a set of candidates, {x1 , x2 , . . . , xk }, to partition the data. In this situation, the algorithm is actually using a multiple comparison procedure to decide whether a decision tree should be extended. More speciﬁcally, it is testing for ∆(T0 , Txmax ) > α instead of ∆(T0 , Tx ) > α. As the number of alternatives, k, increases, so does our chance of ﬁnding ∆(T0 , Txmax ) > α. Unless the gain function ∆ or threshold α is modiﬁed to account for k, the algorithm may inadvertently add spurious nodes to the model, which leads to model overﬁtting. This eﬀect becomes more pronounced when the number of training records from which xmax is chosen is small, because the variance of ∆(T0 , Txmax ) is high when fewer examples are available for training. As a result, the probability of ﬁnding ∆(T0 , Txmax ) > α increases when there are very few training records. This often happens when the decision tree grows deeper, which in turn reduces the number of records covered by the nodes and increases the likelihood of adding unnecessary nodes into the tree. Failure to compensate for the large number of alternatives or the small number of training records will therefore lead to model overﬁtting.
4.4.4
Estimation of Generalization Errors
Although the primary reason for overﬁtting is still a subject of debate, it is generally agreed that the complexity of a model has an impact on model overﬁtting, as was illustrated in Figure 4.23. The question is, how do we
179
Chapter 4
Classiﬁcation
determine the right model complexity? The ideal complexity is that of a model that produces the lowest generalization error. The problem is that the learning algorithm has access only to the training set during model building (see Figure 4.3). It has no knowledge of the test set, and thus, does not know how well the tree will perform on records it has never seen before. The best it can do is to estimate the generalization error of the induced tree. This section presents several methods for doing the estimation. Using Resubstitution Estimate The resubstitution estimate approach assumes that the training set is a good representation of the overall data. Consequently, the training error, otherwise known as resubstitution error, can be used to provide an optimistic estimate for the generalization error. Under this assumption, a decision tree induction algorithm simply selects the model that produces the lowest training error rate as its ﬁnal model. However, the training error is usually a poor estimate of generalization error. Example 4.1. Consider the binary decision trees shown in Figure 4.27. Assume that both trees are generated from the same training data and both make their classiﬁcation decisions at each leaf node according to the majority class. Note that the left tree, TL , is more complex because it expands some of the leaf nodes in the right tree, TR . The training error rate for the left tree is e(TL ) = 4/24 = 0.167, while the training error rate for the right tree is
+: 3 –: 0
+: 3 –: 1
+: 2 –: 1
+: 0 –: 2
+: 1 –: 2
Decision Tree, TL
+: 5 –: 2
+: 3 –: 1
+: 1 –: 4
+: 3 –: 0
+: 3 –: 6
+: 0 –: 5 Decision Tree, TR
Figure 4.27. Example of two decision trees generated from the same training data.
180
4.4
Model Overﬁtting
e(TR ) = 6/24 = 0.25. Based on their resubstitution estimate, the left tree is considered better than the right tree. Incorporating Model Complexity As previously noted, the chance for model overﬁtting increases as the model becomes more complex. For this reason, we should prefer simpler models, a strategy that agrees with a wellknown principle known as Occam’s razor or the principle of parsimony: Deﬁnition 4.2. Occam’s Razor: Given two models with the same generalization errors, the simpler model is preferred over the more complex model. Occam’s razor is intuitive because the additional components in a complex model stand a greater chance of being ﬁtted purely by chance. In the words of Einstein, “Everything should be made as simple as possible, but not simpler.” Next, we present two methods for incorporating model complexity into the evaluation of classiﬁcation models. Pessimistic Error Estimate The ﬁrst approach explicitly computes generalization error as the sum of training error and a penalty term for model complexity. The resulting generalization error can be considered its pessimistic error estimate. For instance, let n(t) be the number of training records classiﬁed by node t and e(t) be the number of misclassiﬁed records. The pessimistic error estimate of a decision tree T , eg (T ), can be computed as follows: k eg (T ) =
i=1 [e(ti ) + Ω(ti )] k i=1 n(ti )
=
e(T ) + Ω(T ) , Nt
where k is the number of leaf nodes, e(T ) is the overall training error of the decision tree, Nt is the number of training records, and Ω(ti ) is the penalty term associated with each node ti . Example 4.2. Consider the binary decision trees shown in Figure 4.27. If the penalty term is equal to 0.5, then the pessimistic error estimate for the left tree is 7.5 4 + 7 × 0.5 = = 0.3125 eg (TL ) = 24 24 and the pessimistic error estimate for the right tree is eg (TR ) =
8 6 + 4 × 0.5 = = 0.3333. 24 24
181
Chapter 4
Classiﬁcation A? Yes
No
0
B? B2
C?
A
X X1 X2 X3 X4 ... Xn
B1
y 1 0 0 1 ... 1
1
C1
C2
0
1
Labeled
B
X X1 X2 X3 X4 ... Xn
y ? ? ? ? ... ?
Unlabeled
Figure 4.28. The minimum description length (MDL) principle.
Thus, the left tree has a better pessimistic error rate than the right tree. For binary trees, a penalty term of 0.5 means a node should always be expanded into its two child nodes as long as it improves the classiﬁcation of at least one training record because expanding a node, which is equivalent to adding 0.5 to the overall error, is less costly than committing one training error. If Ω(t) = 1 for all the nodes t, the pessimistic error estimate for the left tree is eg (TL ) = 11/24 = 0.458, while the pessimistic error estimate for the right tree is eg (TR ) = 10/24 = 0.417. The right tree therefore has a better pessimistic error rate than the left tree. Thus, a node should not be expanded into its child nodes unless it reduces the misclassiﬁcation error for more than one training record. Minimum Description Length Principle Another way to incorporate model complexity is based on an informationtheoretic approach known as the minimum description length or MDL principle. To illustrate this principle, consider the example shown in Figure 4.28. In this example, both A and B are given a set of records with known attribute values x. In addition, person A knows the exact class label for each record, while person B knows none of this information. B can obtain the classiﬁcation of each record by requesting that A transmits the class labels sequentially. Such a message would require Θ(n) bits of information, where n is the total number of records. Alternatively, A may decide to build a classiﬁcation model that summarizes the relationship between x and y. The model can be encoded in a compact
182
4.4
Model Overﬁtting
form before being transmitted to B. If the model is 100% accurate, then the cost of transmission is equivalent to the cost of encoding the model. Otherwise, A must also transmit information about which record is classiﬁed incorrectly by the model. Thus, the overall cost of transmission is Cost(model, data) = Cost(model) + Cost(datamodel),
(4.9)
where the ﬁrst term on the righthand side is the cost of encoding the model, while the second term represents the cost of encoding the mislabeled records. According to the MDL principle, we should seek a model that minimizes the overall cost function. An example showing how to compute the total description length of a decision tree is given by Exercise 9 on page 202. Estimating Statistical Bounds The generalization error can also be estimated as a statistical correction to the training error. Since generalization error tends to be larger than training error, the statistical correction is usually computed as an upper bound to the training error, taking into account the number of training records that reach a particular leaf node. For instance, in the C4.5 decision tree algorithm, the number of errors committed by each leaf node is assumed to follow a binomial distribution. To compute its generalization error, we must determine the upper bound limit to the observed training error, as illustrated in the next example. Example 4.3. Consider the leftmost branch of the binary decision trees shown in Figure 4.27. Observe that the leftmost leaf node of TR has been expanded into two child nodes in TL . Before splitting, the error rate of the node is 2/7 = 0.286. By approximating a binomial distribution with a normal distribution, the following upper bound of the error rate e can be derived:
e+ eupper (N, e, α) =
2 zα/2 2N
+ zα/2 1+
e(1−e) N 2 zα/2
+
2 zα/2
4N 2
,
(4.10)
N
where α is the conﬁdence level, zα/2 is the standardized value from a standard normal distribution, and N is the total number of training records used to compute e. By replacing α = 25%, N = 7, and e = 2/7, the upper bound for the error rate is eupper (7, 2/7, 0.25) = 0.503, which corresponds to 7 × 0.503 = 3.521 errors. If we expand the node into its child nodes as shown in TL , the training error rates for the child nodes are 1/4 = 0.250 and 1/3 = 0.333,
183
Chapter 4
Classiﬁcation
respectively. Using Equation 4.10, the upper bounds of these error rates are eupper (4, 1/4, 0.25) = 0.537 and eupper (3, 1/3, 0.25) = 0.650, respectively. The overall training error of the child nodes is 4 × 0.537 + 3 × 0.650 = 4.098, which is larger than the estimated error for the corresponding node in TR . Using a Validation Set In this approach, instead of using the training set to estimate the generalization error, the original training data is divided into two smaller subsets. One of the subsets is used for training, while the other, known as the validation set, is used for estimating the generalization error. Typically, twothirds of the training set is reserved for model building, while the remaining onethird is used for error estimation. This approach is typically used with classiﬁcation techniques that can be parameterized to obtain models with diﬀerent levels of complexity. The complexity of the best model can be estimated by adjusting the parameter of the learning algorithm (e.g., the pruning level of a decision tree) until the empirical model produced by the learning algorithm attains the lowest error rate on the validation set. Although this approach provides a better way for estimating how well the model performs on previously unseen records, less data is available for training.
4.4.5
Handling Overﬁtting in Decision Tree Induction
In the previous section, we described several methods for estimating the generalization error of a classiﬁcation model. Having a reliable estimate of generalization error allows the learning algorithm to search for an accurate model without overﬁtting the training data. This section presents two strategies for avoiding model overﬁtting in the context of decision tree induction. Prepruning (Early Stopping Rule) In this approach, the treegrowing algorithm is halted before generating a fully grown tree that perfectly ﬁts the entire training data. To do this, a more restrictive stopping condition must be used; e.g., stop expanding a leaf node when the observed gain in impurity measure (or improvement in the estimated generalization error) falls below a certain threshold. The advantage of this approach is that it avoids generating overly complex subtrees that overﬁt the training data. Nevertheless, it is diﬃcult to choose the right threshold for early termination. Too high of a threshold will result in underﬁtted models, while a threshold that is set too low may not be suﬃcient to overcome the model overﬁtting problem. Furthermore,
184
4.4
Model Overﬁtting
Decision Tree: depth = 1:  breadth> 7 : class 1  breadth 1:  MultiAgent = 0:   depth > 2: class 0   depth 1:  MultiAgent = 0: class 0  MultiAgent = 1:   totalPages 81: class 1
Subtree Replacement
Figure 4.29. Postpruning of the decision tree for Web robot detection.
even if no signiﬁcant gain is obtained using one of the existing attribute test conditions, subsequent splitting may result in better subtrees. Postpruning In this approach, the decision tree is initially grown to its maximum size. This is followed by a treepruning step, which proceeds to trim the fully grown tree in a bottomup fashion. Trimming can be done by replacing a subtree with (1) a new leaf node whose class label is determined from the majority class of records aﬃliated with the subtree, or (2) the most frequently used branch of the subtree. The treepruning step terminates when no further improvement is observed. Postpruning tends to give better results than prepruning because it makes pruning decisions based on a fully grown tree, unlike prepruning, which can suﬀer from premature termination of the treegrowing process. However, for postpruning, the additional computations needed to grow the full tree may be wasted when the subtree is pruned. Figure 4.29 illustrates the simpliﬁed decision tree model for the Web robot detection example given in Section 4.3.6. Notice that the subtrees rooted at
185
Chapter 4
Classiﬁcation
depth = 1 have been replaced by one of the branches involving the attribute ImagePages. This approach is also known as subtree raising. The depth > 1 and MultiAgent = 0 subtree has been replaced by a leaf node assigned to class 0. This approach is known as subtree replacement. The subtree for depth > 1 and MultiAgent = 1 remains intact.
4.5
Evaluating the Performance of a Classiﬁer
Section 4.4.4 described several methods for estimating the generalization error of a model during training. The estimated error helps the learning algorithm to do model selection; i.e., to ﬁnd a model of the right complexity that is not susceptible to overﬁtting. Once the model has been constructed, it can be applied to the test set to predict the class labels of previously unseen records. It is often useful to measure the performance of the model on the test set because such a measure provides an unbiased estimate of its generalization error. The accuracy or error rate computed from the test set can also be used to compare the relative performance of diﬀerent classiﬁers on the same domain. However, in order to do this, the class labels of the test records must be known. This section reviews some of the methods commonly used to evaluate the performance of a classiﬁer.
4.5.1
Holdout Method
In the holdout method, the original data with labeled examples is partitioned into two disjoint sets, called the training and the test sets, respectively. A classiﬁcation model is then induced from the training set and its performance is evaluated on the test set. The proportion of data reserved for training and for testing is typically at the discretion of the analysts (e.g., 5050 or twothirds for training and onethird for testing). The accuracy of the classiﬁer can be estimated based on the accuracy of the induced model on the test set. The holdout method has several wellknown limitations. First, fewer labeled examples are available for training because some of the records are withheld for testing. As a result, the induced model may not be as good as when all the labeled examples are used for training. Second, the model may be highly dependent on the composition of the training and test sets. The smaller the training set size, the larger the variance of the model. On the other hand, if the training set is too large, then the estimated accuracy computed from the smaller test set is less reliable. Such an estimate is said to have a wide conﬁdence interval. Finally, the training and test sets are no longer independent
186
4.5
Evaluating the Performance of a Classiﬁer
of each other. Because the training and test sets are subsets of the original data, a class that is overrepresented in one subset will be underrepresented in the other, and vice versa.
4.5.2
Random Subsampling
The holdout method can be repeated several times to improve the estimation of a classiﬁer’s performance. This approach is known as random subsampling. during the ith iteration. The overall accuracy Let acci be the model accuracy k is given by accsub = i=1 acci /k. Random subsampling still encounters some of the problems associated with the holdout method because it does not utilize as much data as possible for training. It also has no control over the number of times each record is used for testing and training. Consequently, some records might be used for training more often than others.
4.5.3
CrossValidation
An alternative to random subsampling is crossvalidation. In this approach, each record is used the same number of times for training and exactly once for testing. To illustrate this method, suppose we partition the data into two equalsized subsets. First, we choose one of the subsets for training and the other for testing. We then swap the roles of the subsets so that the previous training set becomes the test set and vice versa. This approach is called a twofold crossvalidation. The total error is obtained by summing up the errors for both runs. In this example, each record is used exactly once for training and once for testing. The kfold crossvalidation method generalizes this approach by segmenting the data into k equalsized partitions. During each run, one of the partitions is chosen for testing, while the rest of them are used for training. This procedure is repeated k times so that each partition is used for testing exactly once. Again, the total error is found by summing up the errors for all k runs. A special case of the kfold crossvalidation method sets k = N , the size of the data set. In this socalled leaveoneout approach, each test set contains only one record. This approach has the advantage of utilizing as much data as possible for training. In addition, the test sets are mutually exclusive and they eﬀectively cover the entire data set. The drawback of this approach is that it is computationally expensive to repeat the procedure N times. Furthermore, since each test set contains only one record, the variance of the estimated performance metric tends to be high.
187
Chapter 4
4.5.4
Classiﬁcation
Bootstrap
The methods presented so far assume that the training records are sampled without replacement. As a result, there are no duplicate records in the training and test sets. In the bootstrap approach, the training records are sampled with replacement; i.e., a record already chosen for training is put back into the original pool of records so that it is equally likely to be redrawn. If the original data has N records, it can be shown that, on average, a bootstrap sample of size N contains about 63.2% of the records in the original data. This approximation follows from the fact that the probability a record is chosen by a bootstrap sample is 1 − (1 − 1/N )N . When N is suﬃciently large, the probability asymptotically approaches 1 − e−1 = 0.632. Records that are not included in the bootstrap sample become part of the test set. The model induced from the training set is then applied to the test set to obtain an estimate of the accuracy of the bootstrap sample, i . The sampling procedure is then repeated b times to generate b bootstrap samples. There are several variations to the bootstrap sampling approach in terms of how the overall accuracy of the classiﬁer is computed. One of the more widely used approaches is the .632 bootstrap, which computes the overall accuracy by combining the accuracies of each bootstrap sample ( i ) with the accuracy computed from a training set that contains all the labeled examples in the original data (accs ): 1 (0.632 × i + 0.368 × accs ). b b
Accuracy, accboot =
(4.11)
i=1
4.6
Methods for Comparing Classiﬁers
It is often useful to compare the performance of diﬀerent classiﬁers to determine which classiﬁer works better on a given data set. However, depending on the size of the data, the observed diﬀerence in accuracy between two classiﬁers may not be statistically signiﬁcant. This section examines some of the statistical tests available to compare the performance of diﬀerent models and classiﬁers. For illustrative purposes, consider a pair of classiﬁcation models, MA and MB . Suppose MA achieves 85% accuracy when evaluated on a test set containing 30 records, while MB achieves 75% accuracy on a diﬀerent test set containing 5000 records. Based on this information, is MA a better model than MB ?
188
4.6
Methods for Comparing Classiﬁers
The preceding example raises two key questions regarding the statistical signiﬁcance of the performance metrics: 1. Although MA has a higher accuracy than MB , it was tested on a smaller test set. How much conﬁdence can we place on the accuracy for MA ? 2. Is it possible to explain the diﬀerence in accuracy as a result of variations in the composition of the test sets? The ﬁrst question relates to the issue of estimating the conﬁdence interval of a given model accuracy. The second question relates to the issue of testing the statistical signiﬁcance of the observed deviation. These issues are investigated in the remainder of this section.
4.6.1
Estimating a Conﬁdence Interval for Accuracy
To determine the conﬁdence interval, we need to establish the probability distribution that governs the accuracy measure. This section describes an approach for deriving the conﬁdence interval by modeling the classiﬁcation task as a binomial experiment. Following is a list of characteristics of a binomial experiment: 1. The experiment consists of N independent trials, where each trial has two possible outcomes: success or failure. 2. The probability of success, p, in each trial is constant. An example of a binomial experiment is counting the number of heads that turn up when a coin is ﬂipped N times. If X is the number of successes observed in N trials, then the probability that X takes a particular value is given by a binomial distribution with mean N p and variance N p(1 − p): N v P (X = v) = p (1 − p)N −v . p For example, if the coin is fair (p = 0.5) and is ﬂipped ﬁfty times, then the probability that the head shows up 20 times is 50 P (X = 20) = 0.520 (1 − 0.5)30 = 0.0419. 20 If the experiment is repeated many times, then the average number of heads expected to show up is 50 × 0.5 = 25, while its variance is 50 × 0.5 × 0.5 = 12.5.
189
Chapter 4
Classiﬁcation
The task of predicting the class labels of test records can also be considered as a binomial experiment. Given a test set that contains N records, let X be the number of records correctly predicted by a model and p be the true accuracy of the model. By modeling the prediction task as a binomial experiment, X has a binomial distribution with mean N p and variance N p(1 − p). It can be shown that the empirical accuracy, acc = X/N , also has a binomial distribution with mean p and variance p(1 − p)/N (see Exercise 12). Although the binomial distribution can be used to estimate the conﬁdence interval for acc, it is often approximated by a normal distribution when N is suﬃciently large. Based on the normal distribution, the following conﬁdence interval for acc can be derived: acc − p (4.12) ≤ Z1−α/2 = 1 − α, P − Zα/2 ≤ p(1 − p)/N where Zα/2 and Z1−α/2 are the upper and lower bounds obtained from a standard normal distribution at conﬁdence level (1 − α). Since a standard normal distribution is symmetric around Z = 0, it follows that Zα/2 = Z1−α/2 . Rearranging this inequality leads to the following conﬁdence interval for p:
2 ±Z 2 2 2 × N × acc + Zα/2 α/2 Zα/2 + 4N acc − 4N acc . (4.13) 2 ) 2(N + Zα/2 The following table shows the values of Zα/2 at diﬀerent conﬁdence levels: 1−α Zα/2
0.99 2.58
0.98 2.33
0.95 1.96
0.9 1.65
0.8 1.28
0.7 1.04
0.5 0.67
Example 4.4. Consider a model that has an accuracy of 80% when evaluated on 100 test records. What is the conﬁdence interval for its true accuracy at a 95% conﬁdence level? The conﬁdence level of 95% corresponds to Zα/2 = 1.96 according to the table given above. Inserting this term into Equation 4.13 yields a conﬁdence interval between 71.1% and 86.7%. The following table shows the conﬁdence interval when the number of records, N , increases: N Conﬁdence Interval
20 0.584 − 0.919
50 0.670 − 0.888
100 0.711 − 0.867
500 0.763 − 0.833
1000 0.774 − 0.824
5000 0.789 − 0.811
Note that the conﬁdence interval becomes tighter when N increases.
190
4.6
4.6.2
Methods for Comparing Classiﬁers
Comparing the Performance of Two Models
Consider a pair of models, M1 and M2 , that are evaluated on two independent test sets, D1 and D2 . Let n1 denote the number of records in D1 and n2 denote the number of records in D2 . In addition, suppose the error rate for M1 on D1 is e1 and the error rate for M2 on D2 is e2 . Our goal is to test whether the observed diﬀerence between e1 and e2 is statistically signiﬁcant. Assuming that n1 and n2 are suﬃciently large, the error rates e1 and e2 can be approximated using normal distributions. If the observed diﬀerence in the error rate is denoted as d = e1 − e2 , then d is also normally distributed with mean dt , its true diﬀerence, and variance, σd2 . The variance of d can be computed as follows: σd2 σ d2 =
e1 (1 − e1 ) e2 (1 − e2 ) + , n1 n2
(4.14)
where e1 (1 − e1 )/n1 and e2 (1 − e2 )/n2 are the variances of the error rates. Finally, at the (1 − α)% conﬁdence level, it can be shown that the conﬁdence interval for the true diﬀerence dt is given by the following equation: d . dt = d ± zα/2 σ
(4.15)
Example 4.5. Consider the problem described at the beginning of this section. Model MA has an error rate of e1 = 0.15 when applied to N1 = 30 test records, while model MB has an error rate of e2 = 0.25 when applied to N2 = 5000 test records. The observed diﬀerence in their error rates is d = 0.15 − 0.25 = 0.1. In this example, we are performing a twosided test to check whether dt = 0 or dt = 0. The estimated variance of the observed diﬀerence in error rates can be computed as follows: σ d2 =
0.15(1 − 0.15) 0.25(1 − 0.25) + = 0.0043 30 5000
or σ d = 0.0655. Inserting this value into Equation 4.15, we obtain the following conﬁdence interval for dt at 95% conﬁdence level: dt = 0.1 ± 1.96 × 0.0655 = 0.1 ± 0.128. As the interval spans the value zero, we can conclude that the observed diﬀerence is not statistically signiﬁcant at a 95% conﬁdence level.
191
Chapter 4
Classiﬁcation
At what conﬁdence level can we reject the hypothesis that dt = 0? To do this, we need to determine the value of Zα/2 such that the conﬁdence interval for dt does not span the value zero. We can reverse the preceding computation and look for the value Zα/2 such that d > Zα/2 σ d . Replacing the values of d and σ d gives Zα/2 < 1.527. This value ﬁrst occurs when (1 − α) 0.936 (for a twosided test). The result suggests that the null hypothesis can be rejected at conﬁdence level of 93.6% or lower.
4.6.3
Comparing the Performance of Two Classiﬁers
Suppose we want to compare the performance of two classiﬁers using the kfold crossvalidation approach. Initially, the data set D is divided into k equalsized partitions. We then apply each classiﬁer to construct a model from k − 1 of the partitions and test it on the remaining partition. This step is repeated k times, each time using a diﬀerent partition as the test set. Let Mij denote the model induced by classiﬁcation technique Li during the th j iteration. Note that each pair of models M1j and M2j are tested on the same partition j. Let e1j and e2j be their respective error rates. The diﬀerence between their error rates during the j th fold can be written as dj = e1j − e2j . If k is suﬃciently large, then dj is normally distributed with mean dcv t , which is the true diﬀerence in their error rates, and variance σ cv . Unlike the previous approach, the overall variance in the observed diﬀerences is estimated using the following formula: k 2 j=1 (dj − d) , (4.16) σ d2cv = k(k − 1) where d is the average diﬀerence. For this approach, we need to use a tdistribution to compute the conﬁdence interval for dcv t : dcv dcv . t = d ± t(1−α),k−1 σ The coeﬃcient t(1−α),k−1 is obtained from a probability table with two input parameters, its conﬁdence level (1 − α) and the number of degrees of freedom, k − 1. The probability table for the tdistribution is shown in Table 4.6. Example 4.6. Suppose the estimated diﬀerence in the accuracy of models generated by two classiﬁcation techniques has a mean equal to 0.05 and a standard deviation equal to 0.002. If the accuracy is estimated using a 30fold crossvalidation approach, then at a 95% conﬁdence level, the true accuracy diﬀerence is (4.17) dcv t = 0.05 ± 2.04 × 0.002.
192
4.7
Bibliographic Notes
Table 4.6. Probability table for tdistribution.
k−1 1 2 4 9 14 19 24 29
0.99 3.08 1.89 1.53 1.38 1.34 1.33 1.32 1.31
0.98 6.31 2.92 2.13 1.83 1.76 1.73 1.71 1.70
(1 − α) 0.95 12.7 4.30 2.78 2.26 2.14 2.09 2.06 2.04
0.9 31.8 6.96 3.75 2.82 2.62 2.54 2.49 2.46
0.8 63.7 9.92 4.60 3.25 2.98 2.86 2.80 2.76
Since the conﬁdence interval does not span the value zero, the observed difference between the techniques is statistically signiﬁcant.
4.7
Bibliographic Notes
Early classiﬁcation systems were developed to organize a large collection of objects. For example, the Dewey Decimal and Library of Congress classiﬁcation systems were designed to catalog and index the vast number of library books. The categories are typically identiﬁed in a manual fashion, with the help of domain experts. Automated classiﬁcation has been a subject of intensive research for many years. The study of classiﬁcation in classical statistics is sometimes known as discriminant analysis, where the objective is to predict the group membership of an object based on a set of predictor variables. A wellknown classical method is Fisher’s linear discriminant analysis [117], which seeks to ﬁnd a linear projection of the data that produces the greatest discrimination between objects that belong to diﬀerent classes. Many pattern recognition problems also require the discrimination of objects from diﬀerent classes. Examples include speech recognition, handwritten character identiﬁcation, and image classiﬁcation. Readers who are interested in the application of classiﬁcation techniques for pattern recognition can refer to the survey articles by Jain et al. [122] and Kulkarni et al. [128] or classic pattern recognition books by Bishop [107], Duda et al. [114], and Fukunaga [118]. The subject of classiﬁcation is also a major research topic in the ﬁelds of neural networks, statistical learning, and machine learning. An indepth treat
193
Chapter 4
Classiﬁcation
ment of various classiﬁcation techniques is given in the books by Cherkassky and Mulier [112], Hastie et al. [120], Michie et al. [133], and Mitchell [136]. An overview of decision tree induction algorithms can be found in the survey articles by Buntine [110], Moret [137], Murthy [138], and Safavian et al. [147]. Examples of some wellknown decision tree algorithms include CART [108], ID3 [143], C4.5 [145], and CHAID [125]. Both ID3 and C4.5 employ the entropy measure as their splitting function. An indepth discussion of the C4.5 decision tree algorithm is given by Quinlan [145]. Besides explaining the methodology for decision tree growing and tree pruning, Quinlan [145] also described how the algorithm can be modiﬁed to handle data sets with missing values. The CART algorithm was developed by Breiman et al. [108] and uses the Gini index as its splitting function. CHAID [125] uses the statistical χ2 test to determine the best split during the treegrowing process. The decision tree algorithm presented in this chapter assumes that the splitting condition is speciﬁed one attribute at a time. An oblique decision tree can use multiple attributes to form the attribute test condition in the internal nodes [121, 152]. Breiman et al. [108] provide an option for using linear combinations of attributes in their CART implementation. Other approaches for inducing oblique decision trees were proposed by Heath et al. [121], Murthy et al. [139], Cant´ uPaz and Kamath [111], and Utgoﬀ and Brodley [152]. Although oblique decision trees help to improve the expressiveness of a decision tree representation, learning the appropriate test condition at each node is computationally challenging. Another way to improve the expressiveness of a decision tree without using oblique decision trees is to apply a method known as constructive induction [132]. This method simpliﬁes the task of learning complex splitting functions by creating compound features from the original attributes. Besides the topdown approach, other strategies for growing a decision tree include the bottomup approach by Landeweerd et al. [130] and Pattipati and Alexandridis [142], as well as the bidirectional approach by Kim and Landgrebe [126]. Schuermann and Doster [150] and Wang and Suen [154] proposed using a soft splitting criterion to address the data fragmentation problem. In this approach, each record is assigned to diﬀerent branches of the decision tree with diﬀerent probabilities. Model overﬁtting is an important issue that must be addressed to ensure that a decision tree classiﬁer performs equally well on previously unknown records. The model overﬁtting problem has been investigated by many authors including Breiman et al. [108], Schaﬀer [148], Mingers [135], and Jensen and Cohen [123]. While the presence of noise is often regarded as one of the
194
Bibliography primary reasons for overﬁtting [135, 140], Jensen and Cohen [123] argued that overﬁtting is the result of using incorrect hypothesis tests in a multiple comparison procedure. Schapire [149] deﬁned generalization error as “the probability of misclassifying a new example” and test error as “the fraction of mistakes on a newly sampled test set.” Generalization error can therefore be considered as the expected test error of a classiﬁer. Generalization error may sometimes refer to the true error [136] of a model, i.e., its expected error for randomly drawn data points from the same population distribution where the training set is sampled. These deﬁnitions are in fact equivalent if both the training and test sets are gathered from the same population distribution, which is often the case in many data mining and machine learning applications. The Occam’s razor principle is often attributed to the philosopher William of Occam. Domingos [113] cautioned against the pitfall of misinterpreting Occam’s razor as comparing models with similar training errors, instead of generalization errors. A survey on decision treepruning methods to avoid overﬁtting is given by Breslow and Aha [109] and Esposito et al. [116]. Some of the typical pruning methods include reduced error pruning [144], pessimistic error pruning [144], minimum error pruning [141], critical value pruning [134], costcomplexity pruning [108], and errorbased pruning [145]. Quinlan and Rivest proposed using the minimum description length principle for decision tree pruning in [146]. Kohavi [127] had performed an extensive empirical study to compare the performance metrics obtained using diﬀerent estimation methods such as random subsampling, bootstrapping, and kfold crossvalidation. Their results suggest that the best estimation method is based on the tenfold stratiﬁed crossvalidation. Efron and Tibshirani [115] provided a theoretical and empirical comparison between crossvalidation and a bootstrap method known as the 632+ rule. Current techniques such as C4.5 require that the entire training data set ﬁt into main memory. There has been considerable eﬀort to develop parallel and scalable versions of decision tree induction algorithms. Some of the proposed algorithms include SLIQ by Mehta et al. [131], SPRINT by Shafer et al. [151], CMP by Wang and Zaniolo [153], CLOUDS by Alsabti et al. [106], RainForest by Gehrke et al. [119], and ScalParC by Joshi et al. [124]. A general survey of parallel algorithms for data mining is available in [129].
195
Chapter 4
Classiﬁcation
Bibliography [106] K. Alsabti, S. Ranka, and V. Singh. CLOUDS: A Decision Tree Classiﬁer for Large Datasets. In Proc. of the 4th Intl. Conf. on Knowledge Discovery and Data Mining, pages 2–8, New York, NY, August 1998. [107] C. M. Bishop. Neural Networks for Pattern Recognition. Oxford University Press, Oxford, U.K., 1995. [108] L. Breiman, J. H. Friedman, R. Olshen, and C. J. Stone. Classification and Regression Trees. Chapman & Hall, New York, 1984. [109] L. A. Breslow and D. W. Aha. Simplifying Decision Trees: A Survey. Knowledge Engineering Review, 12(1):1–40, 1997. [110] W. Buntine. Learning classiﬁcation trees. In Artificial Intelligence Frontiers in Statistics, pages 182–201. Chapman & Hall, London, 1993. [111] E. Cant´ uPaz and C. Kamath. Using evolutionary algorithms to induce oblique decision trees. In Proc. of the Genetic and Evolutionary Computation Conf., pages 1053–1060, San Francisco, CA, 2000. [112] V. Cherkassky and F. Mulier. Learning from Data: Concepts, Theory, and Methods. Wiley Interscience, 1998. [113] P. Domingos. The Role of Occam’s Razor in Knowledge Discovery. Data Mining and Knowledge Discovery, 3(4):409–425, 1999. [114] R. O. Duda, P. E. Hart, and D. G. Stork. Pattern Classification. John Wiley & Sons, Inc., New York, 2nd edition, 2001. [115] B. Efron and R. Tibshirani. Crossvalidation and the Bootstrap: Estimating the Error Rate of a Prediction Rule. Technical report, Stanford University, 1995. [116] F. Esposito, D. Malerba, and G. Semeraro. A Comparative Analysis of Methods for Pruning Decision Trees. IEEE Trans. Pattern Analysis and Machine Intelligence, 19 (5):476–491, May 1997. [117] R. A. Fisher. The use of multiple measurements in taxonomic problems. Annals of Eugenics, 7:179–188, 1936. [118] K. Fukunaga. Introduction to Statistical Pattern Recognition. Academic Press, New York, 1990. [119] J. Gehrke, R. Ramakrishnan, and V. Ganti. RainForest—A Framework for Fast Decision Tree Construction of Large Datasets. Data Mining and Knowledge Discovery, 4 (2/3):127–162, 2000. [120] T. Hastie, R. Tibshirani, and J. H. Friedman. The Elements of Statistical Learning: Data Mining, Inference, Prediction. Springer, New York, 2001. [121] D. Heath, S. Kasif, and S. Salzberg. Induction of Oblique Decision Trees. In Proc. of the 13th Intl. Joint Conf. on Artificial Intelligence, pages 1002–1007, Chambery, France, August 1993. [122] A. K. Jain, R. P. W. Duin, and J. Mao. Statistical Pattern Recognition: A Review. IEEE Tran. Patt. Anal. and Mach. Intellig., 22(1):4–37, 2000. [123] D. Jensen and P. R. Cohen. Multiple Comparisons in Induction Algorithms. Machine Learning, 38(3):309–338, March 2000. [124] M. V. Joshi, G. Karypis, and V. Kumar. ScalParC: A New Scalable and Eﬃcient Parallel Classiﬁcation Algorithm for Mining Large Datasets. In Proc. of 12th Intl. Parallel Processing Symp. (IPPS/SPDP), pages 573–579, Orlando, FL, April 1998. [125] G. V. Kass. An Exploratory Technique for Investigating Large Quantities of Categorical Data. Applied Statistics, 29:119–127, 1980.
196
Bibliography [126] B. Kim and D. Landgrebe. Hierarchical decision classiﬁers in highdimensional and large class data. IEEE Trans. on Geoscience and Remote Sensing, 29(4):518–528, 1991. [127] R. Kohavi. A Study on CrossValidation and Bootstrap for Accuracy Estimation and Model Selection. In Proc. of the 15th Intl. Joint Conf. on Artificial Intelligence, pages 1137–1145, Montreal, Canada, August 1995. [128] S. R. Kulkarni, G. Lugosi, and S. S. Venkatesh. Learning Pattern Classiﬁcation—A Survey. IEEE Tran. Inf. Theory, 44(6):2178–2206, 1998. [129] V. Kumar, M. V. Joshi, E.H. Han, P. N. Tan, and M. Steinbach. High Performance Data Mining. In High Performance Computing for Computational Science (VECPAR 2002), pages 111–125. Springer, 2002. [130] G. Landeweerd, T. Timmers, E. Gersema, M. Bins, and M. Halic. Binary tree versus single level tree classiﬁcation of white blood cells. Pattern Recognition, 16:571–577, 1983. [131] M. Mehta, R. Agrawal, and J. Rissanen. SLIQ: A Fast Scalable Classiﬁer for Data Mining. In Proc. of the 5th Intl. Conf. on Extending Database Technology, pages 18–32, Avignon, France, March 1996. [132] R. S. Michalski. A theory and methodology of inductive learning. Artificial Intelligence, 20:111–116, 1983. [133] D. Michie, D. J. Spiegelhalter, and C. C. Taylor. Machine Learning, Neural and Statistical Classification. Ellis Horwood, Upper Saddle River, NJ, 1994. [134] J. Mingers. Expert Systems—Rule Induction with Statistical Data. J Operational Research Society, 38:39–47, 1987. [135] J. Mingers. An empirical comparison of pruning methods for decision tree induction. Machine Learning, 4:227–243, 1989. [136] T. Mitchell. Machine Learning. McGrawHill, Boston, MA, 1997. [137] B. M. E. Moret. Decision Trees and Diagrams. Computing Surveys, 14(4):593–623, 1982. [138] S. K. Murthy. Automatic Construction of Decision Trees from Data: A MultiDisciplinary Survey. Data Mining and Knowledge Discovery, 2(4):345–389, 1998. [139] S. K. Murthy, S. Kasif, and S. Salzberg. A system for induction of oblique decision trees. J of Artificial Intelligence Research, 2:1–33, 1994. [140] T. Niblett. Constructing decision trees in noisy domains. In Proc. of the 2nd European Working Session on Learning, pages 67–78, Bled, Yugoslavia, May 1987. [141] T. Niblett and I. Bratko. Learning Decision Rules in Noisy Domains. In Research and Development in Expert Systems III, Cambridge, 1986. Cambridge University Press. [142] K. R. Pattipati and M. G. Alexandridis. Application of heuristic search and information theory to sequential fault diagnosis. IEEE Trans. on Systems, Man, and Cybernetics, 20(4):872–887, 1990. [143] J. R. Quinlan. Discovering rules by induction from large collection of examples. In D. Michie, editor, Expert Systems in the Micro Electronic Age. Edinburgh University Press, Edinburgh, UK, 1979. [144] J. R. Quinlan. Simplifying Decision Trees. Intl. J. ManMachine Studies, 27:221–234, 1987. [145] J. R. Quinlan. C4.5: Programs for Machine Learning. MorganKaufmann Publishers, San Mateo, CA, 1993. [146] J. R. Quinlan and R. L. Rivest. Inferring Decision Trees Using the Minimum Description Length Principle. Information and Computation, 80(3):227–248, 1989.
197
Chapter 4
Classiﬁcation
[147] S. R. Safavian and D. Landgrebe. A Survey of Decision Tree Classiﬁer Methodology. IEEE Trans. Systems, Man and Cybernetics, 22:660–674, May/June 1998. [148] C. Schaﬀer. Overﬁtting avoidence as bias. Machine Learning, 10:153–178, 1993. [149] R. E. Schapire. The Boosting Approach to Machine Learning: An Overview. In MSRI Workshop on Nonlinear Estimation and Classification, 2002. [150] J. Schuermann and W. Doster. A decisiontheoretic approach in hierarchical classiﬁer design. Pattern Recognition, 17:359–369, 1984. [151] J. C. Shafer, R. Agrawal, and M. Mehta. SPRINT: A Scalable Parallel Classiﬁer for Data Mining. In Proc. of the 22nd VLDB Conf., pages 544–555, Bombay, India, September 1996. [152] P. E. Utgoﬀ and C. E. Brodley. An incremental method for ﬁnding multivariate splits for decision trees. In Proc. of the 7th Intl. Conf. on Machine Learning, pages 58–65, Austin, TX, June 1990. [153] H. Wang and C. Zaniolo. CMP: A Fast Decision Tree Classiﬁer Using Multivariate Predictions. In Proc. of the 16th Intl. Conf. on Data Engineering, pages 449–460, San Diego, CA, March 2000. [154] Q. R. Wang and C. Y. Suen. Large tree classiﬁer with heuristic search and global training. IEEE Trans. on Pattern Analysis and Machine Intelligence, 9(1):91–102, 1987.
4.8
Exercises
1. Draw the full decision tree for the parity function of four Boolean attributes, A, B, C, and D. Is it possible to simplify the tree? 2. Consider the training examples shown in Table 4.7 for a binary classiﬁcation problem. (a) Compute the Gini index for the overall collection of training examples. (b) Compute the Gini index for the Customer ID attribute. (c) Compute the Gini index for the Gender attribute. (d) Compute the Gini index for the Car Type attribute using multiway split. (e) Compute the Gini index for the Shirt Size attribute using multiway split. (f) Which attribute is better, Gender, Car Type, or Shirt Size? (g) Explain why Customer ID should not be used as the attribute test condition even though it has the lowest Gini. 3. Consider the training examples shown in Table 4.8 for a binary classiﬁcation problem. (a) What is the entropy of this collection of training examples with respect to the positive class?
198
4.8 Table 4.7. Data set for Exercise 2. Customer ID Gender Car Type Shirt Size 1 M Family Small 2 M Sports Medium 3 M Sports Medium 4 M Sports Large 5 M Sports Extra Large 6 M Sports Extra Large 7 F Sports Small 8 F Sports Small 9 F Sports Medium 10 F Luxury Large 11 M Family Large 12 M Family Extra Large 13 M Family Medium 14 M Luxury Extra Large 15 F Luxury Small 16 F Luxury Small 17 F Luxury Medium 18 F Luxury Medium 19 F Luxury Medium 20 F Luxury Large
Exercises
Class C0 C0 C0 C0 C0 C0 C0 C0 C0 C0 C1 C1 C1 C1 C1 C1 C1 C1 C1 C1
Table 4.8. Data set for Exercise 3. Instance a1 a2 a3 Target Class 1 T T 1.0 + 2 T T 6.0 + 3 T F 5.0 − 4 F F 4.0 + 5 F T 7.0 − 6 F T 3.0 − 7 F F 8.0 − 8 T F 7.0 + 9 F T 5.0 −
(b) What are the information gains of a1 and a2 relative to these training examples? (c) For a3 , which is a continuous attribute, compute the information gain for every possible split.
199
Chapter 4
Classiﬁcation
(d) What is the best split (among a1 , a2 , and a3 ) according to the information gain? (e) What is the best split (between a1 and a2 ) according to the classiﬁcation error rate? (f) What is the best split (between a1 and a2 ) according to the Gini index? 4. Show that the entropy of a node never increases after splitting it into smaller successor nodes. 5. Consider the following data set for a binary class problem. A T T T T T F F F T T
B F T T F T F F F T F
Class Label + + + − + − − − − −
(a) Calculate the information gain when splitting on A and B. Which attribute would the decision tree induction algorithm choose? (b) Calculate the gain in the Gini index when splitting on A and B. Which attribute would the decision tree induction algorithm choose? (c) Figure 4.13 shows that entropy and the Gini index are both monotonously increasing on the range [0, 0.5] and they are both monotonously decreasing on the range [0.5, 1]. Is it possible that information gain and the gain in the Gini index favor diﬀerent attributes? Explain. 6. Consider the following set of training examples. X 0 0 0 0 1 1 1 1
200
Y 0 0 1 1 0 0 1 1
Z 0 1 0 1 0 1 0 1
No. of Class C1 Examples 5 0 10 45 10 25 5 0
No. of Class C2 Examples 40 15 5 0 5 0 20 15
4.8
Exercises
(a) Compute a twolevel decision tree using the greedy approach described in this chapter. Use the classiﬁcation error rate as the criterion for splitting. What is the overall error rate of the induced tree? (b) Repeat part (a) using X as the ﬁrst splitting attribute and then choose the best remaining attribute for splitting at each of the two successor nodes. What is the error rate of the induced tree? (c) Compare the results of parts (a) and (b). Comment on the suitability of the greedy heuristic used for splitting attribute selection. 7. The following table summarizes a data set with three attributes A, B, C and two class labels +, −. Build a twolevel decision tree.
A T F T F T F T F
B T T F F T T F F
C T T T T F F F F
Number of Instances + − 5 0 0 20 20 0 0 5 0 0 25 0 0 0 0 25
(a) According to the classiﬁcation error rate, which attribute would be chosen as the ﬁrst splitting attribute? For each attribute, show the contingency table and the gains in classiﬁcation error rate. (b) Repeat for the two children of the root node. (c) How many instances are misclassiﬁed by the resulting decision tree? (d) Repeat parts (a), (b), and (c) using C as the splitting attribute. (e) Use the results in parts (c) and (d) to conclude about the greedy nature of the decision tree induction algorithm. 8. Consider the decision tree shown in Figure 4.30. (a) Compute the generalization error rate of the tree using the optimistic approach. (b) Compute the generalization error rate of the tree using the pessimistic approach. (For simplicity, use the strategy of adding a factor of 0.5 to each leaf node.) (c) Compute the generalization error rate of the tree using the validation set shown above. This approach is known as reduced error pruning.
201
Chapter 4
Classiﬁcation
A 0
1
B 0
C 1
0
_
+
+
1 _
Training: Instance 1 2 3 4 5 6 7 8 9 10
A 0 0 0 0 1 1 1 1 1 1
B 0 0 1 1 0 0 1 0 1 1
C 0 1 0 1 0 0 0 1 0 0
Class + + + – + + – + – –
Validation: Instance 11 12 13 14 15
A 0 0 1 1 1
B 0 1 1 0 0
C 0 1 0 1 0
Class + + + – +
Figure 4.30. Decision tree and data sets for Exercise 8.
9. Consider the decision trees shown in Figure 4.31. Assume they are generated from a data set that contains 16 binary attributes and 3 classes, C1 , C2 , and C3 .
C1 C1 C2
C2
C3 C1
(a) Decision tree with 7 errors
C2
(b) Decision tree with 4 errors
Figure 4.31. Decision trees for Exercise 9.
202
C3
4.8
Exercises
Compute the total description length of each decision tree according to the minimum description length principle. • The total description length of a tree is given by: Cost(tree, data) = Cost(tree) + Cost(datatree). • Each internal node of the tree is encoded by the ID of the splitting attribute. If there are m attributes, the cost of encoding each attribute is log2 m bits. • Each leaf is encoded using the ID of the class it is associated with. If there are k classes, the cost of encoding a class is log2 k bits. • Cost(tree) is the cost of encoding all the nodes in the tree. To simplify the computation, you can assume that the total cost of the tree is obtained by adding up the costs of encoding each internal node and each leaf node. • Cost(datatree) is encoded using the classiﬁcation errors the tree commits on the training set. Each error is encoded by log2 n bits, where n is the total number of training instances. Which decision tree is better, according to the MDL principle? 10. While the .632 bootstrap approach is useful for obtaining a reliable estimate of model accuracy, it has a known limitation [127]. Consider a twoclass problem, where there are equal number of positive and negative examples in the data. Suppose the class labels for the examples are generated randomly. The classiﬁer used is an unpruned decision tree (i.e., a perfect memorizer). Determine the accuracy of the classiﬁer using each of the following methods. (a) The holdout method, where twothirds of the data are used for training and the remaining onethird are used for testing. (b) Tenfold crossvalidation. (c) The .632 bootstrap method. (d) From the results in parts (a), (b), and (c), which method provides a more reliable evaluation of the classiﬁer’s accuracy? 11. Consider the following approach for testing whether a classiﬁer A beats another classiﬁer B. Let N be the size of a given data set, pA be the accuracy of classiﬁer A, pB be the accuracy of classiﬁer B, and p = (pA + pB )/2 be the average accuracy for both classiﬁers. To test whether classiﬁer A is signiﬁcantly better than B, the following Zstatistic is used: pA − p B Z=
. 2p(1−p) N
Classiﬁer A is assumed to be better than classiﬁer B if Z > 1.96.
203
Chapter 4
Classiﬁcation
Table 4.9 compares the accuracies of three diﬀerent classiﬁers, decision tree classiﬁers, na¨ıve Bayes classiﬁers, and support vector machines, on various data sets. (The latter two classiﬁers are described in Chapter 5.)
Table 4.9. Comparing the accuracy of various classification methods. Data Set Anneal Australia Auto Breast Cleve Credit Diabetes German Glass Heart Hepatitis Horse Ionosphere Iris Labor Led7 Lymphography Pima Sonar Tictactoe Vehicle Wine Zoo
Size (N ) 898 690 205 699 303 690 768 1000 214 270 155 368 351 150 57 3200 148 768 208 958 846 178 101
Decision Tree (%) 92.09 85.51 81.95 95.14 76.24 85.80 72.40 70.90 67.29 80.00 81.94 85.33 89.17 94.67 78.95 73.34 77.03 74.35 78.85 83.72 71.04 94.38 93.07
na¨ıve Bayes (%) 79.62 76.81 58.05 95.99 83.50 77.54 75.91 74.70 48.59 84.07 83.23 78.80 82.34 95.33 94.74 73.16 83.11 76.04 69.71 70.04 45.04 96.63 93.07
Support vector machine (%) 87.19 84.78 70.73 96.42 84.49 85.07 76.82 74.40 59.81 83.70 87.10 82.61 88.89 96.00 92.98 73.56 86.49 76.95 76.92 98.33 74.94 98.88 96.04
Summarize the performance of the classiﬁers given in Table 4.9 using the following 3 × 3 table: winlossdraw Decision tree Na¨ıve Bayes Support vector machine
Decision tree
Na¨ıve Bayes
Support vector machine
0  0  23 0  0  23 0  0  23
Each cell in the table contains the number of wins, losses, and draws when comparing the classiﬁer in a given row to the classiﬁer in a given column.
204
4.8
Exercises
12. Let X be a binomial random variable with mean N p and variance N p(1 − p). Show that the ratio X/N also has a binomial distribution with mean p and variance p(1 − p)/N .
205
206
5 Classification: Alternative Techniques The previous chapter described a simple, yet quite eﬀective, classiﬁcation technique known as decision tree induction. Issues such as model overﬁtting and classiﬁer evaluation were also discussed in great detail. This chapter presents alternative techniques for building classiﬁcation models—from simple techniques such as rulebased and nearestneighbor classiﬁers to more advanced techniques such as support vector machines and ensemble methods. Other key issues such as the class imbalance and multiclass problems are also discussed at the end of the chapter.
5.1
RuleBased Classiﬁer
A rulebased classiﬁer is a technique for classifying records using a collection of “if . . .then. . .” rules. Table 5.1 shows an example of a model generated by a rulebased classiﬁer for the vertebrate classiﬁcation problem. The rules for the model are represented in a disjunctive normal form, R = (r1 ∨r2 ∨. . . rk ), where R is known as the rule set and ri ’s are the classiﬁcation rules or disjuncts. Table 5.1. Example of a rule set for the vertebrate classification problem. r1 : r2 : r3 : r4 : r5 :
(Gives Birth = no) ∧ (Aerial Creature = yes) −→ Birds (Gives Birth = no) ∧ (Aquatic Creature = yes) −→ Fishes (Gives Birth = yes) ∧ (Body Temperature = warmblooded) −→ Mammals (Gives Birth = no) ∧ (Aerial Creature = no) −→ Reptiles (Aquatic Creature = semi) −→ Amphibians
From Chapter 5 of Introduction to Data Mining, First Edition. PangNing Tan, Michael Steinbach, Vipin Kumar. Copyright © 2006 by Pearson Education, Inc. All rights reserved.
207
Chapter 5
Classiﬁcation: Alternative Techniques
Each classiﬁcation rule can be expressed in the following way: ri : (Conditioni ) −→ yi .
(5.1)
The lefthand side of the rule is called the rule antecedent or precondition. It contains a conjunction of attribute tests: Conditioni = (A1 op v1 ) ∧ (A2 op v2 ) ∧ . . . (Ak op vk ),
(5.2)
where (Aj , vj ) is an attributevalue pair and op is a logical operator chosen from the set {=, =, , ≤, ≥}. Each attribute test (Aj op vj ) is known as a conjunct. The righthand side of the rule is called the rule consequent, which contains the predicted class yi . A rule r covers a record x if the precondition of r matches the attributes of x. r is also said to be ﬁred or triggered whenever it covers a given record. For an illustration, consider the rule r1 given in Table 5.1 and the following attributes for two vertebrates: hawk and grizzly bear. Name hawk grizzly bear
Body Temperature warmblooded warmblooded
Skin Cover feather fur
Gives Birth no yes
Aquatic Creature no no
Aerial Creature yes no
Has Legs yes yes
Hibernates no yes
r1 covers the ﬁrst vertebrate because its precondition is satisﬁed by the hawk’s attributes. The rule does not cover the second vertebrate because grizzly bears give birth to their young and cannot ﬂy, thus violating the precondition of r1 . The quality of a classiﬁcation rule can be evaluated using measures such as coverage and accuracy. Given a data set D and a classiﬁcation rule r : A −→ y, the coverage of the rule is deﬁned as the fraction of records in D that trigger the rule r. On the other hand, its accuracy or conﬁdence factor is deﬁned as the fraction of records triggered by r whose class labels are equal to y. The formal deﬁnitions of these measures are Coverage(r) = Accuracy(r) =
A D A ∩ y , A
(5.3)
where A is the number of records that satisfy the rule antecedent, A ∩ y is the number of records that satisfy both the antecedent and consequent, and D is the total number of records.
208
5.1
RuleBased Classiﬁer
Table 5.2. The vertebrate data set. Name human python salmon whale frog komodo dragon bat pigeon cat guppy alligator penguin porcupine eel salamander
Body Temperature warmblooded coldblooded coldblooded warmblooded coldblooded coldblooded
Skin Cover hair scales scales hair none scales
Gives Birth yes no no yes no no
Aquatic Creature no no yes yes semi no
Aerial Creature no no no no no no
Has Legs yes no no no yes yes
Hibernates no yes no no yes no
Class Label Mammals Reptiles Fishes Mammals Amphibians Reptiles
warmblooded warmblooded warmblooded coldblooded coldblooded warmblooded warmblooded coldblooded coldblooded
hair feathers fur scales scales feathers quills scales none
yes no yes yes no no yes no no
no no no yes semi semi no yes semi
yes yes no no no no no no no
yes yes yes no yes yes yes no yes
yes no no no no no yes no yes
Mammals Birds Mammals Fishes Reptiles Birds Mammals Fishes Amphibians
Example 5.1. Consider the data set shown in Table 5.2. The rule (Gives Birth = yes) ∧ (Body Temperature = warmblooded) −→ Mammals has a coverage of 33% since ﬁve of the ﬁfteen records support the rule antecedent. The rule accuracy is 100% because all ﬁve vertebrates covered by the rule are mammals.
5.1.1
How a RuleBased Classiﬁer Works
A rulebased classiﬁer classiﬁes a test record based on the rule triggered by the record. To illustrate how a rulebased classiﬁer works, consider the rule set shown in Table 5.1 and the following vertebrates:
Name lemur turtle dogﬁsh shark
Body Temperature warmblooded coldblooded coldblooded
Skin Cover fur scales scales
Gives Birth yes no yes
Aquatic Creature no semi yes
Aerial Creature no no no
Has Legs yes yes no
Hibernates yes no no
• The ﬁrst vertebrate, which is a lemur, is warmblooded and gives birth to its young. It triggers the rule r3 , and thus, is classiﬁed as a mammal.
209
Chapter 5
Classiﬁcation: Alternative Techniques
• The second vertebrate, which is a turtle, triggers the rules r4 and r5 . Since the classes predicted by the rules are contradictory (reptiles versus amphibians), their conﬂicting classes must be resolved. • None of the rules are applicable to a dogﬁsh shark. In this case, we need to ensure that the classiﬁer can still make a reliable prediction even though a test record is not covered by any rule. The previous example illustrates two important properties of the rule set generated by a rulebased classiﬁer. Mutually Exclusive Rules The rules in a rule set R are mutually exclusive if no two rules in R are triggered by the same record. This property ensures that every record is covered by at most one rule in R. An example of a mutually exclusive rule set is shown in Table 5.3. Exhaustive Rules A rule set R has exhaustive coverage if there is a rule for each combination of attribute values. This property ensures that every record is covered by at least one rule in R. Assuming that Body Temperature and Gives Birth are binary variables, the rule set shown in Table 5.3 has exhaustive coverage. Table 5.3. Example of a mutually exclusive and exhaustive rule set. r1 : (Body Temperature = coldblooded) −→ Nonmammals r2 : (Body Temperature = warmblooded) ∧ (Gives Birth = yes) −→ Mammals r3 : (Body Temperature = warmblooded) ∧ (Gives Birth = no) −→ Nonmammals
Together, these properties ensure that every record is covered by exactly one rule. Unfortunately, many rulebased classiﬁers, including the one shown in Table 5.1, do not have such properties. If the rule set is not exhaustive, then a default rule, rd : () −→ yd , must be added to cover the remaining cases. A default rule has an empty antecedent and is triggered when all other rules have failed. yd is known as the default class and is typically assigned to the majority class of training records not covered by the existing rules. If the rule set is not mutually exclusive, then a record can be covered by several rules, some of which may predict conﬂicting classes. There are two ways to overcome this problem.
210
5.1
RuleBased Classiﬁer
Ordered Rules In this approach, the rules in a rule set are ordered in decreasing order of their priority, which can be deﬁned in many ways (e.g., based on accuracy, coverage, total description length, or the order in which the rules are generated). An ordered rule set is also known as a decision list. When a test record is presented, it is classiﬁed by the highestranked rule that covers the record. This avoids the problem of having conﬂicting classes predicted by multiple classiﬁcation rules. Unordered Rules This approach allows a test record to trigger multiple classiﬁcation rules and considers the consequent of each rule as a vote for a particular class. The votes are then tallied to determine the class label of the test record. The record is usually assigned to the class that receives the highest number of votes. In some cases, the vote may be weighted by the rule’s accuracy. Using unordered rules to build a rulebased classiﬁer has both advantages and disadvantages. Unordered rules are less susceptible to errors caused by the wrong rule being selected to classify a test record (unlike classiﬁers based on ordered rules, which are sensitive to the choice of ruleordering criteria). Model building is also less expensive because the rules do not have to be kept in sorted order. Nevertheless, classifying a test record can be quite an expensive task because the attributes of the test record must be compared against the precondition of every rule in the rule set. In the remainder of this section, we will focus on rulebased classiﬁers that use ordered rules.
5.1.2
RuleOrdering Schemes
Rule ordering can be implemented on a rulebyrule basis or on a classbyclass basis. The diﬀerence between these schemes is illustrated in Figure 5.1. RuleBased Ordering Scheme This approach orders the individual rules by some rule quality measure. This ordering scheme ensures that every test record is classiﬁed by the “best” rule covering it. A potential drawback of this scheme is that lowerranked rules are much harder to interpret because they assume the negation of the rules preceding them. For example, the fourth rule shown in Figure 5.1 for rulebased ordering, Aquatic Creature = semi −→ Amphibians, has the following interpretation: If the vertebrate does not have any feathers or cannot ﬂy, and is coldblooded and semiaquatic, then it is an amphibian.
211
Chapter 5
Classiﬁcation: Alternative Techniques
RuleBased Ordering
ClassBased Ordering
(Skin Cover=feathers, Aerial Creature=yes) ==> Birds
(Skin Cover=feathers, Aerial Creature=yes) ==> Birds
(Body temperature=warmblooded, Gives Birth=yes) ==> Mammals
(Body temperature=warmblooded, Gives Birth=no) ==> Birds
(Body temperature=warmblooded, Gives Birth=no) ==> Birds
(Body temperature=warmblooded, Gives Birth=yes) ==> Mammals
(Aquatic Creature=semi)) ==> Amphibians
(Aquatic Creature=semi)) ==> Amphibians
(Skin Cover=scales, Aquatic Creature=no) ==> Reptiles
(Skin Cover=none) ==> Amphibians
(Skin Cover=scales, Aquatic Creature=yes) ==> Fishes (Skin Cover=none) ==> Amphibians
(Skin Cover=scales, Aquatic Creature=no) ==> Reptiles (Skin Cover=scales, Aquatic Creature=yes) ==> Fishes
Figure 5.1. Comparison between rulebased and classbased ordering schemes.
The additional conditions (that the vertebrate does not have any feathers or cannot ﬂy, and is coldblooded) are due to the fact that the vertebrate does not satisfy the ﬁrst three rules. If the number of rules is large, interpreting the meaning of the rules residing near the bottom of the list can be a cumbersome task. ClassBased Ordering Scheme In this approach, rules that belong to the same class appear together in the rule set R. The rules are then collectively sorted on the basis of their class information. The relative ordering among the rules from the same class is not important; as long as one of the rules ﬁres, the class will be assigned to the test record. This makes rule interpretation slightly easier. However, it is possible for a highquality rule to be overlooked in favor of an inferior rule that happens to predict the higherranked class. Since most of the wellknown rulebased classiﬁers (such as C4.5rules and RIPPER) employ the classbased ordering scheme, the discussion in the remainder of this section focuses mainly on this type of ordering scheme.
5.1.3
How to Build a RuleBased Classiﬁer
To build a rulebased classiﬁer, we need to extract a set of rules that identiﬁes key relationships between the attributes of a data set and the class label.
212
5.1
RuleBased Classiﬁer
There are two broad classes of methods for extracting classiﬁcation rules: (1) direct methods, which extract classiﬁcation rules directly from data, and (2) indirect methods, which extract classiﬁcation rules from other classiﬁcation models, such as decision trees and neural networks. Direct methods partition the attribute space into smaller subspaces so that all the records that belong to a subspace can be classiﬁed using a single classiﬁcation rule. Indirect methods use the classiﬁcation rules to provide a succinct description of more complex classiﬁcation models. Detailed discussions of these methods are presented in Sections 5.1.4 and 5.1.5, respectively.
5.1.4
Direct Methods for Rule Extraction
The sequential covering algorithm is often used to extract rules directly from data. Rules are grown in a greedy fashion based on a certain evaluation measure. The algorithm extracts the rules one class at a time for data sets that contain more than two classes. For the vertebrate classiﬁcation problem, the sequential covering algorithm may generate rules for classifying birds ﬁrst, followed by rules for classifying mammals, amphibians, reptiles, and ﬁnally, ﬁshes (see Figure 5.1). The criterion for deciding which class should be generated ﬁrst depends on a number of factors, such as the class prevalence (i.e., fraction of training records that belong to a particular class) or the cost of misclassifying records from a given class. A summary of the sequential covering algorithm is given in Algorithm 5.1. The algorithm starts with an empty decision list, R. The LearnOneRule function is then used to extract the best rule for class y that covers the current set of training records. During rule extraction, all training records for class y are considered to be positive examples, while those that belong to Algorithm 5.1 Sequential covering algorithm. 1: 2: 3: 4: 5: 6: 7: 8: 9: 10: 11:
Let E be the training records and A be the set of attributevalue pairs, {(Aj , vj )}. Let Yo be an ordered set of classes {y1 , y2 , . . . , yk }. Let R = { } be the initial rule list. for each class y ∈ Yo − {yk } do while stopping condition is not met do r ← LearnOneRule (E, A, y). Remove training records from E that are covered by r. Add r to the bottom of the rule list: R −→ R ∨ r. end while end for Insert the default rule, {} −→ yk , to the bottom of the rule list R.
213
Chapter 5
Classiﬁcation: Alternative Techniques
other classes are considered to be negative examples. A rule is desirable if it covers most of the positive examples and none (or very few) of the negative examples. Once such a rule is found, the training records covered by the rule are eliminated. The new rule is added to the bottom of the decision list R. This procedure is repeated until the stopping criterion is met. The algorithm then proceeds to generate rules for the next class. Figure 5.2 demonstrates how the sequential covering algorithm works for a data set that contains a collection of positive and negative examples. The rule R1, whose coverage is shown in Figure 5.2(b), is extracted ﬁrst because it covers the largest fraction of positive examples. All the training records covered by R1 are subsequently removed and the algorithm proceeds to look for the next best rule, which is R2.
R1
(a) Original Data
R1
(b) Step 1
R1
R2 (c) Step 2
(d) Step 3
Figure 5.2. An example of the sequential covering algorithm.
214
5.1
RuleBased Classiﬁer
LearnOneRule Function The objective of the LearnOneRule function is to extract a classiﬁcation rule that covers many of the positive examples and none (or very few) of the negative examples in the training set. However, ﬁnding an optimal rule is computationally expensive given the exponential size of the search space. The LearnOneRule function addresses the exponential search problem by growing the rules in a greedy fashion. It generates an initial rule r and keeps reﬁning the rule until a certain stopping criterion is met. The rule is then pruned to improve its generalization error. RuleGrowing Strategy There are two common strategies for growing a classiﬁcation rule: generaltospeciﬁc or speciﬁctogeneral. Under the generaltospeciﬁc strategy, an initial rule r : {} −→ y is created, where the lefthand side is an empty set and the righthand side contains the target class. The rule has poor quality because it covers all the examples in the training set. New
{ } => Mammals
Skin Cover = hair => Mammals
Body Temperature = warmblooded => Mammals
...
Body Temperature = warmblooded, Has Legs = yes => Mammals
...
Has Legs = No => Mammals
Body Temperature = warmblooded, Gives Birth = yes => Mammals
(a) Generaltospecific
Body Temperature=warmblooded, Skin Cover=hair, Gives Birth=yes, Aquatic creature=no, Aerial Creature=no Has Legs=yes, Hibernates=no => Mammals
Skin Cover=hair, Gives Birth=yes Aquatic Creature=no, Aerial Creature=no, Has Legs=yes, Hibernates=no => Mammals
...
Body Temperature=warmblooded, Skin Cover=hair, Gives Birth=yes, Aquatic creature=no, Aerial Creature=no Has Legs=yes => Mammals
(b) Specifictogeneral
Figure 5.3. Generaltospecific and specifictogeneral rulegrowing strategies.
215
Chapter 5
Classiﬁcation: Alternative Techniques
conjuncts are subsequently added to improve the rule’s quality. Figure 5.3(a) shows the generaltospeciﬁc rulegrowing strategy for the vertebrate classiﬁcation problem. The conjunct Body Temperature=warmblooded is initially chosen to form the rule antecedent. The algorithm then explores all the possible candidates and greedily chooses the next conjunct, Gives Birth=yes, to be added into the rule antecedent. This process continues until the stopping criterion is met (e.g., when the added conjunct does not improve the quality of the rule). For the speciﬁctogeneral strategy, one of the positive examples is randomly chosen as the initial seed for the rulegrowing process. During the reﬁnement step, the rule is generalized by removing one of its conjuncts so that it can cover more positive examples. Figure 5.3(b) shows the speciﬁctogeneral approach for the vertebrate classiﬁcation problem. Suppose a positive example for mammals is chosen as the initial seed. The initial rule contains the same conjuncts as the attribute values of the seed. To improve its coverage, the rule is generalized by removing the conjunct Hibernate=no. The reﬁnement step is repeated until the stopping criterion is met, e.g., when the rule starts covering negative examples. The previous approaches may produce suboptimal rules because the rules are grown in a greedy fashion. To avoid this problem, a beam search may be used, where k of the best candidate rules are maintained by the algorithm. Each candidate rule is then grown separately by adding (or removing) a conjunct from its antecedent. The quality of the candidates are evaluated and the k best candidates are chosen for the next iteration. Rule Evaluation An evaluation metric is needed to determine which conjunct should be added (or removed) during the rulegrowing process. Accuracy is an obvious choice because it explicitly measures the fraction of training examples classiﬁed correctly by the rule. However, a potential limitation of accuracy is that it does not take into account the rule’s coverage. For example, consider a training set that contains 60 positive examples and 100 negative examples. Suppose we are given the following two candidate rules: Rule r1 : covers 50 positive examples and 5 negative examples, Rule r2 : covers 2 positive examples and no negative examples. The accuracies for r1 and r2 are 90.9% and 100%, respectively. However, r1 is the better rule despite its lower accuracy. The high accuracy for r2 is potentially spurious because the coverage of the rule is too low.
216
5.1
RuleBased Classiﬁer
The following approaches can be used to handle this problem. 1. A statistical test can be used to prune rules that have poor coverage. For example, we may compute the following likelihood ratio statistic:
R=2
k
fi log(fi /ei ),
i=1
where k is the number of classes, fi is the observed frequency of class i examples that are covered by the rule, and ei is the expected frequency of a rule that makes random predictions. Note that R has a chisquare distribution with k − 1 degrees of freedom. A large R value suggests that the number of correct predictions made by the rule is signiﬁcantly larger than that expected by random guessing. For example, since r1 covers 55 examples, the expected frequency for the positive class is e+ = 55×60/160 = 20.625, while the expected frequency for the negative class is e− = 55 × 100/160 = 34.375. Thus, the likelihood ratio for r1 is R(r1 ) = 2 × [50 × log2 (50/20.625) + 5 × log2 (5/34.375)] = 99.9. Similarly, the expected frequencies for r2 are e+ = 2 × 60/160 = 0.75 and e− = 2 × 100/160 = 1.25. The likelihood ratio statistic for r2 is R(r2 ) = 2 × [2 × log2 (2/0.75) + 0 × log2 (0/1.25)] = 5.66. This statistic therefore suggests that r1 is a better rule than r2 . 2. An evaluation metric that takes into account the rule coverage can be used. Consider the following evaluation metrics: Laplace = mestimate =
f+ + 1 , n+k f+ + kp+ , n+k
(5.4) (5.5)
where n is the number of examples covered by the rule, f+ is the number of positive examples covered by the rule, k is the total number of classes, and p+ is the prior probability for the positive class. Note that the mestimate is equivalent to the Laplace measure by choosing p+ = 1/k. Depending on the rule coverage, these measures capture the tradeoﬀ
217
Chapter 5
Classiﬁcation: Alternative Techniques
between rule accuracy and the prior probability of the positive class. If the rule does not cover any training example, then the Laplace measure reduces to 1/k, which is the prior probability of the positive class assuming a uniform class distribution. The mestimate also reduces to the prior probability (p+ ) when n = 0. However, if the rule coverage is large, then both measures asymptotically approach the rule accuracy, f+ /n. Going back to the previous example, the Laplace measure for r1 is 51/57 = 89.47%, which is quite close to its accuracy. Conversely, the Laplace measure for r2 (75%) is signiﬁcantly lower than its accuracy because r2 has a much lower coverage. 3. An evaluation metric that takes into account the support count of the rule can be used. One such metric is the FOIL’s information gain. The support count of a rule corresponds to the number of positive examples covered by the rule. Suppose the rule r : A −→ + covers p0 positive examples and n0 negative examples. After adding a new conjunct B, the extended rule r : A ∧ B −→ + covers p1 positive examples and n1 negative examples. Given this information, the FOIL’s information gain of the extended rule is deﬁned as follows: FOIL’s information gain = p1 ×
log2
p1 p0 . (5.6) − log2 p1 + n 1 p0 + n 0
Since the measure is proportional to p1 and p1 /(p1 + n1 ), it prefers rules that have high support count and accuracy. The FOIL’s information gains for rules r1 and r2 given in the preceding example are 43.12 and 2, respectively. Therefore, r1 is a better rule than r2 . Rule Pruning The rules generated by the LearnOneRule function can be pruned to improve their generalization errors. To determine whether pruning is necessary, we may apply the methods described in Section 4.4 on page 172 to estimate the generalization error of a rule. For example, if the error on validation set decreases after pruning, we should keep the simpliﬁed rule. Another approach is to compare the pessimistic error of the rule before and after pruning (see Section 4.4.4 on page 179). The simpliﬁed rule is retained in place of the original rule if the pessimistic error improves after pruning.
218
5.1
RuleBased Classiﬁer
Rationale for Sequential Covering After a rule is extracted, the sequential covering algorithm must eliminate all the positive and negative examples covered by the rule. The rationale for doing this is given in the next example. R3 R1
+ + + ++ + + + + + + + + + ++ + + + + +  
+ class = +
class = 
R2 +
+
+ + + + + 

Figure 5.4. Elimination of training records by the sequential covering algorithm. R1, R2, and R3 represent regions covered by three different rules.
Figure 5.4 shows three possible rules, R1, R2, and R3, extracted from a data set that contains 29 positive examples and 21 negative examples. The accuracies of R1, R2, and R3 are 12/15 (80%), 7/10 (70%), and 8/12 (66.7%), respectively. R1 is generated ﬁrst because it has the highest accuracy. After generating R1, it is clear that the positive examples covered by the rule must be removed so that the next rule generated by the algorithm is diﬀerent than R1. Next, suppose the algorithm is given the choice of generating either R2 or R3. Even though R2 has higher accuracy than R3, R1 and R3 together cover 18 positive examples and 5 negative examples (resulting in an overall accuracy of 78.3%), whereas R1 and R2 together cover 19 positive examples and 6 negative examples (resulting in an overall accuracy of 76%). The incremental impact of R2 or R3 on accuracy is more evident when the positive and negative examples covered by R1 are removed before computing their accuracies. In particular, if positive examples covered by R1 are not removed, then we may overestimate the eﬀective accuracy of R3, and if negative examples are not removed, then we may underestimate the accuracy of R3. In the latter case, we might end up preferring R2 over R3 even though half of the false positive errors committed by R3 have already been accounted for by the preceding rule, R1.
219
Chapter 5
Classiﬁcation: Alternative Techniques
RIPPER Algorithm To illustrate the direct method, we consider a widely used rule induction algorithm called RIPPER. This algorithm scales almost linearly with the number of training examples and is particularly suited for building models from data sets with imbalanced class distributions. RIPPER also works well with noisy data sets because it uses a validation set to prevent model overﬁtting. For twoclass problems, RIPPER chooses the majority class as its default class and learns the rules for detecting the minority class. For multiclass problems, the classes are ordered according to their frequencies. Let (y1 , y2 , . . . , yc ) be the ordered classes, where y1 is the least frequent class and yc is the most frequent class. During the ﬁrst iteration, instances that belong to y1 are labeled as positive examples, while those that belong to other classes are labeled as negative examples. The sequential covering method is used to generate rules that discriminate between the positive and negative examples. Next, RIPPER extracts rules that distinguish y2 from other remaining classes. This process is repeated until we are left with yc , which is designated as the default class. Rule Growing RIPPER employs a generaltospeciﬁc strategy to grow a rule and the FOIL’s information gain measure to choose the best conjunct to be added into the rule antecedent. It stops adding conjuncts when the rule starts covering negative examples. The new rule is then pruned based on its performance on the validation set. The following metric is computed to determine whether pruning is needed: (p−n)/(p+n), where p (n) is the number of positive (negative) examples in the validation set covered by the rule. This metric is monotonically related to the rule’s accuracy on the validation set. If the metric improves after pruning, then the conjunct is removed. Pruning is done starting from the last conjunct added to the rule. For example, given a rule ABCD −→ y, RIPPER checks whether D should be pruned ﬁrst, followed by CD, BCD, etc. While the original rule covers only positive examples, the pruned rule may cover some of the negative examples in the training set. Building the Rule Set After generating a rule, all the positive and negative examples covered by the rule are eliminated. The rule is then added into the rule set as long as it does not violate the stopping condition, which is based on the minimum description length principle. If the new rule increases the total description length of the rule set by at least d bits, then RIPPER stops adding rules into its rule set (by default, d is chosen to be 64 bits). Another stopping condition used by RIPPER is that the error rate of the rule on the validation set must not exceed 50%.
220
5.1
RuleBased Classiﬁer
RIPPER also performs additional optimization steps to determine whether some of the existing rules in the rule set can be replaced by better alternative rules. Readers who are interested in the details of the optimization method may refer to the reference cited at the end of this chapter.
5.1.5
Indirect Methods for Rule Extraction
This section presents a method for generating a rule set from a decision tree. In principle, every path from the root node to the leaf node of a decision tree can be expressed as a classiﬁcation rule. The test conditions encountered along the path form the conjuncts of the rule antecedent, while the class label at the leaf node is assigned to the rule consequent. Figure 5.5 shows an example of a rule set generated from a decision tree. Notice that the rule set is exhaustive and contains mutually exclusive rules. However, some of the rules can be simpliﬁed as shown in the next example. P No
Yes
Q No

Rule Set
R Yes
+
No
Yes
+
Q No
Yes

+
r1: (P=No,Q=No) ==> r2: (P=No,Q=Yes) ==> + r3: (P=Yes,Q=No) ==> + r4: (P=Yes,R=Yes,Q=No) ==> r5: (P=Yes,R=Yes,Q=Yes) ==> +
Figure 5.5. Converting a decision tree into classification rules.
Example 5.2. Consider the following three rules from Figure 5.5: r2 : (P = No) ∧ (Q = Yes) −→ + r3 : (P = Yes) ∧ (R = No) −→ + r5 : (P = Yes) ∧ (R = Yes) ∧ (Q = Yes) −→ + Observe that the rule set always predicts a positive class when the value of Q is Yes. Therefore, we may simplify the rules as follows: r2 : (Q = Yes) −→ + r3: (P = Yes) ∧ (R = No) −→ +.
221
Chapter 5
Classiﬁcation: Alternative Techniques RuleBased Classifier: (Gives Birth=No, Aerial Creature=Yes) => Birds
Gives Birth? Yes
(Gives Birth=No, Aquatic Creature=Yes) => Fishes (Gives Birth=Yes) => Mammals
No
(Gives Birth=No, Aerial Creature=No, Aquatic Creature=No) => Reptiles ( ) => Amphibians
Aquatic Creature
Mammals Yes
No Semi
Fishes
Aerial Creature
Amphibians Yes Birds
No Reptiles
Figure 5.6. Classification rules extracted from a decision tree for the vertebrate classification problem.
r3 is retained to cover the remaining instances of the positive class. Although the rules obtained after simpliﬁcation are no longer mutually exclusive, they are less complex and are easier to interpret. In the following, we describe an approach used by the C4.5rules algorithm to generate a rule set from a decision tree. Figure 5.6 shows the decision tree and resulting classiﬁcation rules obtained for the data set given in Table 5.2. Rule Generation Classiﬁcation rules are extracted for every path from the root to one of the leaf nodes in the decision tree. Given a classiﬁcation rule r : A −→ y, we consider a simpliﬁed rule, r : A −→ y, where A is obtained by removing one of the conjuncts in A. The simpliﬁed rule with the lowest pessimistic error rate is retained provided its error rate is less than that of the original rule. The rulepruning step is repeated until the pessimistic error of the rule cannot be improved further. Because some of the rules may become identical after pruning, the duplicate rules must be discarded. Rule Ordering After generating the rule set, C4.5rules uses the classbased ordering scheme to order the extracted rules. Rules that predict the same class are grouped together into the same subset. The total description length for each subset is computed, and the classes are arranged in increasing order of their total description length. The class that has the smallest description
222
5.2
NearestNeighbor classiﬁers
length is given the highest priority because it is expected to contain the best set of rules. The total description length for a class is given by Lexception + g × Lmodel , where Lexception is the number of bits needed to encode the misclassiﬁed examples, Lmodel is the number of bits needed to encode the model, and g is a tuning parameter whose default value is 0.5. The tuning parameter depends on the number of redundant attributes present in the model. The value of the tuning parameter is small if the model contains many redundant attributes.
5.1.6
Characteristics of RuleBased Classiﬁers
A rulebased classiﬁer has the following characteristics: • The expressiveness of a rule set is almost equivalent to that of a decision tree because a decision tree can be represented by a set of mutually exclusive and exhaustive rules. Both rulebased and decision tree classiﬁers create rectilinear partitions of the attribute space and assign a class to each partition. Nevertheless, if the rulebased classiﬁer allows multiple rules to be triggered for a given record, then a more complex decision boundary can be constructed. • Rulebased classiﬁers are generally used to produce descriptive models that are easier to interpret, but gives comparable performance to the decision tree classiﬁer. • The classbased ordering approach adopted by many rulebased classiﬁers (such as RIPPER) is well suited for handling data sets with imbalanced class distributions.
5.2
NearestNeighbor classiﬁers
The classiﬁcation framework shown in Figure 4.3 involves a twostep process: (1) an inductive step for constructing a classiﬁcation model from data, and (2) a deductive step for applying the model to test examples. Decision tree and rulebased classiﬁers are examples of eager learners because they are designed to learn a model that maps the input attributes to the class label as soon as the training data becomes available. An opposite strategy would be to delay the process of modeling the training data until it is needed to classify the test examples. Techniques that employ this strategy are known as lazy learners. An example of a lazy learner is the Rote classiﬁer, which memorizes the entire training data and performs classiﬁcation only if the attributes of a test instance match one of the training examples exactly. An obvious drawback of
223
Chapter 5
Classiﬁcation: Alternative Techniques
x
(a) 1nearest neighbor
x
x
(b) 2nearest neighbor
(c) 3nearest neighbor
Figure 5.7. The 1, 2, and 3nearest neighbors of an instance.
this approach is that some test records may not be classiﬁed because they do not match any training example. One way to make this approach more ﬂexible is to ﬁnd all the training examples that are relatively similar to the attributes of the test example. These examples, which are known as nearest neighbors, can be used to determine the class label of the test example. The justiﬁcation for using nearest neighbors is best exempliﬁed by the following saying: “If it walks like a duck, quacks like a duck, and looks like a duck, then it’s probably a duck.” A nearestneighbor classiﬁer represents each example as a data point in a ddimensional space, where d is the number of attributes. Given a test example, we compute its proximity to the rest of the data points in the training set, using one of the proximity measures described in Section 2.4 on page 65. The knearest neighbors of a given example z refer to the k points that are closest to z. Figure 5.7 illustrates the 1, 2, and 3nearest neighbors of a data point located at the center of each circle. The data point is classiﬁed based on the class labels of its neighbors. In the case where the neighbors have more than one label, the data point is assigned to the majority class of its nearest neighbors. In Figure 5.7(a), the 1nearest neighbor of the data point is a negative example. Therefore the data point is assigned to the negative class. If the number of nearest neighbors is three, as shown in Figure 5.7(c), then the neighborhood contains two positive examples and one negative example. Using the majority voting scheme, the data point is assigned to the positive class. In the case where there is a tie between the classes (see Figure 5.7(b)), we may randomly choose one of them to classify the data point. The preceding discussion underscores the importance of choosing the right value for k. If k is too small, then the nearestneighbor classiﬁer may be
224
5.2
NearestNeighbor classiﬁers
x
Figure 5.8. knearest neighbor classification with large k.
susceptible to overﬁtting because of noise in the training data. On the other hand, if k is too large, the nearestneighbor classiﬁer may misclassify the test instance because its list of nearest neighbors may include data points that are located far away from its neighborhood (see Figure 5.8).
5.2.1
Algorithm
A highlevel summary of the nearestneighbor classiﬁcation method is given in Algorithm 5.2. The algorithm computes the distance (or similarity) between each test example z = (x , y ) and all the training examples (x, y) ∈ D to determine its nearestneighbor list, Dz . Such computation can be costly if the number of training examples is large. However, eﬃcient indexing techniques are available to reduce the amount of computations needed to ﬁnd the nearest neighbors of a test example. Algorithm 5.2 The knearest neighbor classiﬁcation algorithm. 1: Let k be the number of nearest neighbors and D be the set of training examples. 2: for each test example z = (x , y ) do 3: Compute d(x , x), the distance between z and every example, (x, y) ∈ D. 4: Select Dz ⊆ D, the set of k closest training examples to z. 5: y = argmax (xi ,yi )∈Dz I(v = yi ) 6: end for
v
225
Chapter 5
Classiﬁcation: Alternative Techniques
Once the nearestneighbor list is obtained, the test example is classiﬁed based on the majority class of its nearest neighbors: Majority Voting: y = argmax v
I(v = yi ),
(5.7)
(xi ,yi )∈Dz
where v is a class label, yi is the class label for one of the nearest neighbors, and I(·) is an indicator function that returns the value 1 if its argument is true and 0 otherwise. In the majority voting approach, every neighbor has the same impact on the classiﬁcation. This makes the algorithm sensitive to the choice of k, as shown in Figure 5.7. One way to reduce the impact of k is to weight the inﬂuence of each nearest neighbor xi according to its distance: wi = 1/d(x , xi )2 . As a result, training examples that are located far away from z have a weaker impact on the classiﬁcation compared to those that are located close to z. Using the distanceweighted voting scheme, the class label can be determined as follows: wi × I(v = yi ). (5.8) DistanceWeighted Voting: y = argmax v
5.2.2
(xi ,yi )∈Dz
Characteristics of NearestNeighbor Classiﬁers
The characteristics of the nearestneighbor classiﬁer are summarized below: • Nearestneighbor classiﬁcation is part of a more general technique known as instancebased learning, which uses speciﬁc training instances to make predictions without having to maintain an abstraction (or model) derived from data. Instancebased learning algorithms require a proximity measure to determine the similarity or distance between instances and a classiﬁcation function that returns the predicted class of a test instance based on its proximity to other instances. • Lazy learners such as nearestneighbor classiﬁers do not require model building. However, classifying a test example can be quite expensive because we need to compute the proximity values individually between the test and training examples. In contrast, eager learners often spend the bulk of their computing resources for model building. Once a model has been built, classifying a test example is extremely fast. • Nearestneighbor classiﬁers make their predictions based on local information, whereas decision tree and rulebased classiﬁers attempt to ﬁnd
226
5.3
Bayesian Classiﬁers
a global model that ﬁts the entire input space. Because the classiﬁcation decisions are made locally, nearestneighbor classiﬁers (with small values of k) are quite susceptible to noise. • Nearestneighbor classiﬁers can produce arbitrarily shaped decision boundaries. Such boundaries provide a more ﬂexible model representation compared to decision tree and rulebased classiﬁers that are often constrained to rectilinear decision boundaries. The decision boundaries of nearestneighbor classiﬁers also have high variability because they depend on the composition of training examples. Increasing the number of nearest neighbors may reduce such variability. • Nearestneighbor classiﬁers can produce wrong predictions unless the appropriate proximity measure and data preprocessing steps are taken. For example, suppose we want to classify a group of people based on attributes such as height (measured in meters) and weight (measured in pounds). The height attribute has a low variability, ranging from 1.5 m to 1.85 m, whereas the weight attribute may vary from 90 lb. to 250 lb. If the scale of the attributes are not taken into consideration, the proximity measure may be dominated by diﬀerences in the weights of a person.
5.3
Bayesian Classiﬁers
In many applications the relationship between the attribute set and the class variable is nondeterministic. In other words, the class label of a test record cannot be predicted with certainty even though its attribute set is identical to some of the training examples. This situation may arise because of noisy data or the presence of certain confounding factors that aﬀect classiﬁcation but are not included in the analysis. For example, consider the task of predicting whether a person is at risk for heart disease based on the person’s diet and workout frequency. Although most people who eat healthily and exercise regularly have less chance of developing heart disease, they may still do so because of other factors such as heredity, excessive smoking, and alcohol abuse. Determining whether a person’s diet is healthy or the workout frequency is suﬃcient is also subject to interpretation, which in turn may introduce uncertainties into the learning problem. This section presents an approach for modeling probabilistic relationships between the attribute set and the class variable. The section begins with an introduction to the Bayes theorem, a statistical principle for combining prior
227
Chapter 5
Classiﬁcation: Alternative Techniques
knowledge of the classes with new evidence gathered from data. The use of the Bayes theorem for solving classiﬁcation problems will be explained, followed by a description of two implementations of Bayesian classiﬁers: na¨ıve Bayes and the Bayesian belief network.
5.3.1
Bayes Theorem
Consider a football game between two rival teams: Team 0 and Team 1. Suppose Team 0 wins 65% of the time and Team 1 wins the remaining matches. Among the games won by Team 0, only 30% of them come from playing on Team 1’s football ﬁeld. On the other hand, 75% of the victories for Team 1 are obtained while playing at home. If Team 1 is to host the next match between the two teams, which team will most likely emerge as the winner?
This question can be answered by using the wellknown Bayes theorem. For completeness, we begin with some basic deﬁnitions from probability theory. Readers who are unfamiliar with concepts in probability may refer to Appendix C for a brief review of this topic. Let X and Y be a pair of random variables. Their joint probability, P (X = x, Y = y), refers to the probability that variable X will take on the value x and variable Y will take on the value y. A conditional probability is the probability that a random variable will take on a particular value given that the outcome for another random variable is known. For example, the conditional probability P (Y = yX = x) refers to the probability that the variable Y will take on the value y, given that the variable X is observed to have the value x. The joint and conditional probabilities for X and Y are related in the following way: P (X, Y ) = P (Y X) × P (X) = P (XY ) × P (Y ). (5.9) Rearranging the last two expressions in Equation 5.9 leads to the following formula, known as the Bayes theorem: P (Y X) =
P (XY )P (Y ) . P (X)
(5.10)
The Bayes theorem can be used to solve the prediction problem stated at the beginning of this section. For notational convenience, let X be the random variable that represents the team hosting the match and Y be the random variable that represents the winner of the match. Both X and Y can
228
5.3
Bayesian Classiﬁers
take on values from the set {0, 1}. We can summarize the information given in the problem as follows: Probability Probability Probability Probability
Team Team Team Team
0 1 1 1
wins is P (Y = 0) = 0.65. wins is P (Y = 1) = 1 − P (Y = 0) = 0.35. hosted the match it won is P (X = 1Y = 1) = 0.75. hosted the match won by Team 0 is P (X = 1Y = 0) = 0.3.
Our objective is to compute P (Y = 1X = 1), which is the conditional probability that Team 1 wins the next match it will be hosting, and compares it against P (Y = 0X = 1). Using the Bayes theorem, we obtain P (Y = 1X = 1) = = = = =
P (X = 1Y = 1) × P (Y = 1) P (X = 1) P (X = 1Y = 1) × P (Y = 1) P (X = 1, Y = 1) + P (X = 1, Y = 0) P (X = 1Y = 1) × P (Y = 1) P (X = 1Y = 1)P (Y = 1) + P (X = 1Y = 0)P (Y = 0) 0.75 × 0.35 0.75 × 0.35 + 0.3 × 0.65 0.5738,
where the law of total probability (see Equation C.5 on page 722) was applied in the second line. Furthermore, P (Y = 0X = 1) = 1 − P (Y = 1X = 1) = 0.4262. Since P (Y = 1X = 1) > P (Y = 0X = 1), Team 1 has a better chance than Team 0 of winning the next match.
5.3.2
Using the Bayes Theorem for Classiﬁcation
Before describing how the Bayes theorem can be used for classiﬁcation, let us formalize the classiﬁcation problem from a statistical perspective. Let X denote the attribute set and Y denote the class variable. If the class variable has a nondeterministic relationship with the attributes, then we can treat X and Y as random variables and capture their relationship probabilistically using P (Y X). This conditional probability is also known as the posterior probability for Y , as opposed to its prior probability, P (Y ). During the training phase, we need to learn the posterior probabilities P (Y X) for every combination of X and Y based on information gathered from the training data. By knowing these probabilities, a test record X can be classiﬁed by ﬁnding the class Y that maximizes the posterior probability,
229
Chapter 5
Classiﬁcation: Alternative Techniques
P (Y X ). To illustrate this approach, consider the task of predicting whether a loan borrower will default on their payments. Figure 5.9 shows a training set with the following attributes: Home Owner, Marital Status, and Annual Income. Loan borrowers who defaulted on their payments are classiﬁed as Yes, while those who repaid their loans are classiﬁed as No.
na
te
bi
Tid 1 2 3 4 5 6 7 8 9 10
Home Owner Yes No No Yes No No Yes No No No
al
us
ic
r go
ry ca
Marital Status Single Married Single Married Divorced Married Divorced Single Married Single
uo tin
n
co
Annual Income 125K 100K 70K 120K 95K 60K 220K 85K 75K 90K
ss
a cl
Defaulted Borrower No No No No Yes No No Yes No Yes
Figure 5.9. Training set for predicting the loan default problem.
Suppose we are given a test record with the following attribute set: X = (Home Owner = No, Marital Status = Married, Annual Income = $120K). To classify the record, we need to compute the posterior probabilities P (YesX) and P (NoX) based on information available in the training data. If P (YesX) > P (NoX), then the record is classiﬁed as Yes; otherwise, it is classiﬁed as No. Estimating the posterior probabilities accurately for every possible combination of class label and attribute value is a diﬃcult problem because it requires a very large training set, even for a moderate number of attributes. The Bayes theorem is useful because it allows us to express the posterior probability in terms of the prior probability P (Y ), the classconditional probability P (XY ), and the evidence, P (X): P (Y X) =
P (XY ) × P (Y ) . P (X)
(5.11)
When comparing the posterior probabilities for diﬀerent values of Y , the denominator term, P (X), is always constant, and thus, can be ignored. The
230
5.3
Bayesian Classiﬁers
prior probability P (Y ) can be easily estimated from the training set by computing the fraction of training records that belong to each class. To estimate the classconditional probabilities P (XY ), we present two implementations of Bayesian classiﬁcation methods: the na¨ıve Bayes classiﬁer and the Bayesian belief network. These implementations are described in Sections 5.3.3 and 5.3.5, respectively.
5.3.3
Na¨ıve Bayes Classiﬁer
A na¨ıve Bayes classiﬁer estimates the classconditional probability by assuming that the attributes are conditionally independent, given the class label y. The conditional independence assumption can be formally stated as follows:
P (XY = y) =
d
P (Xi Y = y),
(5.12)
i=1
where each attribute set X = {X1 , X2 , . . . , Xd } consists of d attributes. Conditional Independence Before delving into the details of how a na¨ıve Bayes classiﬁer works, let us examine the notion of conditional independence. Let X, Y, and Z denote three sets of random variables. The variables in X are said to be conditionally independent of Y, given Z, if the following condition holds: P (XY, Z) = P (XZ).
(5.13)
An example of conditional independence is the relationship between a person’s arm length and his or her reading skills. One might observe that people with longer arms tend to have higher levels of reading skills. This relationship can be explained by the presence of a confounding factor, which is age. A young child tends to have short arms and lacks the reading skills of an adult. If the age of a person is ﬁxed, then the observed relationship between arm length and reading skills disappears. Thus, we can conclude that arm length and reading skills are conditionally independent when the age variable is ﬁxed.
231
Chapter 5
Classiﬁcation: Alternative Techniques
The conditional independence between X and Y can also be written into a form that looks similar to Equation 5.12: P (X, Y, Z) P (Z) P (X, Y, Z) P (Y, Z) × = P (Y, Z) P (Z) = P (XY, Z) × P (YZ)
P (X, YZ) =
= P (XZ) × P (YZ),
(5.14)
where Equation 5.13 was used to obtain the last line of Equation 5.14. How a Na¨ıve Bayes Classiﬁer Works With the conditional independence assumption, instead of computing the classconditional probability for every combination of X, we only have to estimate the conditional probability of each Xi , given Y . The latter approach is more practical because it does not require a very large training set to obtain a good estimate of the probability. To classify a test record, the na¨ıve Bayes classiﬁer computes the posterior probability for each class Y : P (Y X) =
P (Y )
d
i=1 P (Xi Y ) . P (X)
(5.15)
Since P (X) is ﬁxed for every Y , it is suﬃcient to choose the class that maximizes the numerator term, P (Y ) di=1 P (Xi Y ). In the next two subsections, we describe several approaches for estimating the conditional probabilities P (Xi Y ) for categorical and continuous attributes. Estimating Conditional Probabilities for Categorical Attributes For a categorical attribute Xi , the conditional probability P (Xi = xi Y = y) is estimated according to the fraction of training instances in class y that take on a particular attribute value xi . For example, in the training set given in Figure 5.9, three out of the seven people who repaid their loans also own a home. As a result, the conditional probability for P (Home Owner=YesNo) is equal to 3/7. Similarly, the conditional probability for defaulted borrowers who are single is given by P (Marital Status = SingleYes) = 2/3.
232
5.3
Bayesian Classiﬁers
Estimating Conditional Probabilities for Continuous Attributes There are two ways to estimate the classconditional probabilities for continuous attributes in na¨ıve Bayes classiﬁers: 1. We can discretize each continuous attribute and then replace the continuous attribute value with its corresponding discrete interval. This approach transforms the continuous attributes into ordinal attributes. The conditional probability P (Xi Y = y) is estimated by computing the fraction of training records belonging to class y that falls within the corresponding interval for Xi . The estimation error depends on the discretization strategy (as described in Section 2.3.6 on page 57), as well as the number of discrete intervals. If the number of intervals is too large, there are too few training records in each interval to provide a reliable estimate for P (Xi Y ). On the other hand, if the number of intervals is too small, then some intervals may aggregate records from diﬀerent classes and we may miss the correct decision boundary. 2. We can assume a certain form of probability distribution for the continuous variable and estimate the parameters of the distribution using the training data. A Gaussian distribution is usually chosen to represent the classconditional probability for continuous attributes. The distribution is characterized by two parameters, its mean, µ, and variance, σ 2 . For each class yj , the classconditional probability for attribute Xi is − 1 P (Xi = xi Y = yj ) = √ exp 2πσij
(xi −µij )2 2σ 2 ij
.
(5.16)
The parameter µij can be estimated based on the sample mean of Xi 2 can (x) for all training records that belong to the class yj . Similarly, σij be estimated from the sample variance (s2 ) of such training records. For example, consider the annual income attribute shown in Figure 5.9. The sample mean and variance for this attribute with respect to the class No are 125 + 100 + 70 + . . . + 75 = 110 7 (125 − 110)2 + (100 − 110)2 + . . . + (75 − 110)2 = 2975 s2 = 7(6) √ s = 2975 = 54.54. x=
233
Chapter 5
Classiﬁcation: Alternative Techniques
Given a test record with taxable income equal to $120K, we can compute its classconditional probability as follows: P (Income=120No) = √
(120−110)2 1 exp− 2×2975 = 0.0072. 2π(54.54)
Note that the preceding interpretation of classconditional probability is somewhat misleading. The righthand side of Equation 5.16 corresponds to a probability density function, f (Xi ; µij , σij ). Since the function is continuous, the probability that the random variable Xi takes a particular value is zero. Instead, we should compute the conditional probability that Xi lies within some interval, xi and xi + , where is a small constant: P (xi ≤ Xi ≤ xi + Y = yj ) =
xi +
f (Xi ; µij , σij )dXi xi
≈ f (xi ; µij , σij ) × .
(5.17)
Since appears as a constant multiplicative factor for each class, it cancels out when we normalize the posterior probability for P (Y X). Therefore, we can still apply Equation 5.16 to approximate the classconditional probability P (Xi Y ). Example of the Na¨ıve Bayes Classiﬁer Consider the data set shown in Figure 5.10(a). We can compute the classconditional probability for each categorical attribute, along with the sample mean and variance for the continuous attribute using the methodology described in the previous subsections. These probabilities are summarized in Figure 5.10(b). To predict the class label of a test record X = (Home Owner=No, Marital Status = Married, Income = $120K), we need to compute the posterior probabilities P (NoX) and P (YesX). Recall from our earlier discussion that these posterior probabilities can be estimated by computing the product between the prior probability P (Y ) and the classconditional probabilities i P (Xi Y ), which corresponds to the numerator of the righthand side term in Equation 5.15. The prior probabilities of each class can be estimated by calculating the fraction of training records that belong to each class. Since there are three records that belong to the class Yes and seven records that belong to the class
234
5.3
Tid 1 2 3 4 5 6 7 8 9 10
Home Owner Yes No No Yes No No Yes No No No
Marital Status Single Married Single Married Divorced Married Divorced Single Married Single
Annual Income 125K 100K 70K 120K 95K 60K 220K 85K 75K 90K
Defaulted Borrower No No No No Yes No No Yes No Yes
Bayesian Classiﬁers
P(Home Owner=YesNo) = 3/7 P(Home Owner=NoNo) = 4/7 P(Home Owner=YesYes) = 0 P(Home Owner=NoYes) = 1 P(Marital Status=SingleNo) = 2/7 P(Marital Status=DivorcedNo) = 1/7 P(Marital Status=MarriedNo) = 4/7 P(Marital Status=SingleYes) = 2/3 P(Marital Status=DivorcedYes) = 1/3 P(Marital Status=MarriedYes) = 0 For Annual Income: If class=No: sample mean=110 sample variance=2975 If class=Yes: sample mean=90 sample variance=25
(a)
(b)
Figure 5.10. The na¨ıve Bayes classifier for the loan classification problem.
No, P (Yes) = 0.3 and P (No) = 0.7. Using the information provided in Figure 5.10(b), the classconditional probabilities can be computed as follows: P (XNo) = P (Home Owner = NoNo) × P (Status = MarriedNo) × P (Annual Income = $120KNo) = 4/7 × 4/7 × 0.0072 = 0.0024. P (XYes) = P (Home Owner = NoYes) × P (Status = MarriedYes) × P (Annual Income = $120KYes) = 1 × 0 × 1.2 × 10−9 = 0. Putting them together, the posterior probability for class No is P (NoX) = α × 7/10 × 0.0024 = 0.0016α, where α = 1/P (X) is a constant term. Using a similar approach, we can show that the posterior probability for class Yes is zero because its classconditional probability is zero. Since P (NoX) > P (YesX), the record is classiﬁed as No.
235
Chapter 5
Classiﬁcation: Alternative Techniques
Mestimate of Conditional Probability The preceding example illustrates a potential problem with estimating posterior probabilities from training data. If the classconditional probability for one of the attributes is zero, then the overall posterior probability for the class vanishes. This approach of estimating classconditional probabilities using simple fractions may seem too brittle, especially when there are few training examples available and the number of attributes is large. In a more extreme case, if the training examples do not cover many of the attribute values, we may not be able to classify some of the test records. For example, if P (Marital Status = DivorcedNo) is zero instead of 1/7, then a record with attribute set X = (Home Owner = Yes, Marital Status = Divorced, Income = $120K) has the following classconditional probabilities: P (XNo) = 3/7 × 0 × 0.0072 = 0. P (XYes) = 0 × 1/3 × 1.2 × 10−9 = 0. The na¨ıve Bayes classiﬁer will not be able to classify the record. This problem can be addressed by using the mestimate approach for estimating the conditional probabilities: P (xi yj ) =
nc + mp , n+m
(5.18)
where n is the total number of instances from class yj , nc is the number of training examples from class yj that take on the value xi , m is a parameter known as the equivalent sample size, and p is a userspeciﬁed parameter. If there is no training set available (i.e., n = 0), then P (xi yj ) = p. Therefore p can be regarded as the prior probability of observing the attribute value xi among records with class yj . The equivalent sample size determines the tradeoﬀ between the prior probability p and the observed probability nc /n. In the example given in the previous section, the conditional probability P (Status = MarriedYes) = 0 because none of the training records for the class has the particular attribute value. Using the mestimate approach with m = 3 and p = 1/3, the conditional probability is no longer zero: P (Marital Status = MarriedYes) = (0 + 3 × 1/3)/(3 + 3) = 1/6.
236
5.3
Bayesian Classiﬁers
If we assume p = 1/3 for all attributes of class Yes and p = 2/3 for all attributes of class No, then P (XNo) = P (Home Owner = NoNo) × P (Status = MarriedNo) × P (Annual Income = $120KNo) = 6/10 × 6/10 × 0.0072 = 0.0026. P (XYes) = P (Home Owner = NoYes) × P (Status = MarriedYes) × P (Annual Income = $120KYes) = 4/6 × 1/6 × 1.2 × 10−9 = 1.3 × 10−10 . The posterior probability for class No is P (NoX) = α × 7/10 × 0.0026 = 0.0018α, while the posterior probability for class Yes is P (YesX) = α × 3/10 × 1.3 × 10−10 = 4.0 × 10−11 α. Although the classiﬁcation decision has not changed, the mestimate approach generally provides a more robust way for estimating probabilities when the number of training examples is small. Characteristics of Na¨ıve Bayes Classiﬁers Na¨ıve Bayes classiﬁers generally have the following characteristics: • They are robust to isolated noise points because such points are averaged out when estimating conditional probabilities from data. Na¨ıve Bayes classiﬁers can also handle missing values by ignoring the example during model building and classiﬁcation. • They are robust to irrelevant attributes. If Xi is an irrelevant attribute, then P (Xi Y ) becomes almost uniformly distributed. The classconditional probability for Xi has no impact on the overall computation of the posterior probability. • Correlated attributes can degrade the performance of na¨ıve Bayes classiﬁers because the conditional independence assumption no longer holds for such attributes. For example, consider the following probabilities: P (A = 0Y = 0) = 0.4, P (A = 1Y = 0) = 0.6, P (A = 0Y = 1) = 0.6, P (A = 1Y = 1) = 0.4, where A is a binary attribute and Y is a binary class variable. Suppose there is another binary attribute B that is perfectly correlated with A
237
Chapter 5
Classiﬁcation: Alternative Techniques
when Y = 0, but is independent of A when Y = 1. For simplicity, assume that the classconditional probabilities for B are the same as for A. Given a record with attributes A = 0, B = 0, we can compute its posterior probabilities as follows: P (Y = 0A = 0, B = 0) = =
P (Y = 1A = 0, B = 0) = =
P (A = 0Y = 0)P (B = 0Y = 0)P (Y = 0) P (A = 0, B = 0) 0.16 × P (Y = 0) . P (A = 0, B = 0) P (A = 0Y = 1)P (B = 0Y = 1)P (Y = 1) P (A = 0, B = 0) 0.36 × P (Y = 1) . P (A = 0, B = 0)
If P (Y = 0) = P (Y = 1), then the na¨ıve Bayes classiﬁer would assign the record to class 1. However, the truth is, P (A = 0, B = 0Y = 0) = P (A = 0Y = 0) = 0.4, because A and B are perfectly correlated when Y = 0. As a result, the posterior probability for Y = 0 is P (Y = 0A = 0, B = 0) = =
P (A = 0, B = 0Y = 0)P (Y = 0) P (A = 0, B = 0) 0.4 × P (Y = 0) , P (A = 0, B = 0)
which is larger than that for Y = 1. The record should have been classiﬁed as class 0.
5.3.4
Bayes Error Rate
Suppose we know the true probability distribution that governs P (XY ). The Bayesian classiﬁcation method allows us to determine the ideal decision boundary for the classiﬁcation task, as illustrated in the following example. Example 5.3. Consider the task of identifying alligators and crocodiles based on their respective lengths. The average length of an adult crocodile is about 15 feet, while the average length of an adult alligator is about 12 feet. Assuming
238
5.3
Bayesian Classiﬁers
0.2 0.18 Crocodile
Alligator 0.16 0.14
P(xy)
0.12 0.1 0.08 0.06 0.04 0.02 0 5
10
15
20
Length, x
Figure 5.11. Comparing the likelihood functions of a crocodile and an alligator.
that their length x follows a Gaussian distribution with a standard deviation equal to 2 feet, we can express their classconditional probabilities as follows: P (XCrocodile) = P (XAlligator) =
1 exp − 2π · 2 1 √ exp − 2π · 2 √
1 X − 15 2 2 2 1 X − 12 2 2 2
(5.19) (5.20)
Figure 5.11 shows a comparison between the classconditional probabilities for a crocodile and an alligator. Assuming that their prior probabilities are the same, the ideal decision boundary is located at some length x ˆ such that P (X = x ˆCrocodile) = P (X = x ˆAlligator). Using Equations 5.19 and 5.20, we obtain
x ˆ − 15 2
2
=
x ˆ − 12 2
2 ,
which can be solved to yield x ˆ = 13.5. The decision boundary for this example is located halfway between the two means.
239
Chapter 5
Classiﬁcation: Alternative Techniques D
A
B
y C
C
X1 A
(a)
X2
X3
X4
...
Xd
B (b)
(c)
Figure 5.12. Representing probabilistic relationships using directed acyclic graphs.
When the prior probabilities are diﬀerent, the decision boundary shifts toward the class with lower prior probability (see Exercise 10 on page 319). Furthermore, the minimum error rate attainable by any classiﬁer on the given data can also be computed. The ideal decision boundary in the preceding example classiﬁes all creatures whose lengths are less than x ˆ as alligators and those whose lengths are greater than x ˆ as crocodiles. The error rate of the classiﬁer is given by the sum of the area under the posterior probability curve for crocodiles (from length 0 to x ˆ) and the area under the posterior probability curve for alligators (from x ˆ to ∞):
x ˆ
P (CrocodileX)dX +
Error = 0
∞
P (AlligatorX)dX. x ˆ
The total error rate is known as the Bayes error rate.
5.3.5
Bayesian Belief Networks
The conditional independence assumption made by na¨ıve Bayes classiﬁers may seem too rigid, especially for classiﬁcation problems in which the attributes are somewhat correlated. This section presents a more ﬂexible approach for modeling the classconditional probabilities P (XY ). Instead of requiring all the attributes to be conditionally independent given the class, this approach allows us to specify which pair of attributes are conditionally independent. We begin with a discussion on how to represent and build such a probabilistic model, followed by an example of how to make inferences from the model.
240
5.3
Bayesian Classiﬁers
Model Representation A Bayesian belief network (BBN), or simply, Bayesian network, provides a graphical representation of the probabilistic relationships among a set of random variables. There are two key elements of a Bayesian network: 1. A directed acyclic graph (dag) encoding the dependence relationships among a set of variables. 2. A probability table associating each node to its immediate parent nodes. Consider three random variables, A, B, and C, in which A and B are independent variables and each has a direct inﬂuence on a third variable, C. The relationships among the variables can be summarized into the directed acyclic graph shown in Figure 5.12(a). Each node in the graph represents a variable, and each arc asserts the dependence relationship between the pair of variables. If there is a directed arc from X to Y , then X is the parent of Y and Y is the child of X. Furthermore, if there is a directed path in the network from X to Z, then X is an ancestor of Z, while Z is a descendant of X. For example, in the diagram shown in Figure 5.12(b), A is a descendant of D and D is an ancestor of B. Both B and D are also nondescendants of A. An important property of the Bayesian network can be stated as follows: Property 1 (Conditional Independence). A node in a Bayesian network is conditionally independent of its nondescendants, if its parents are known. In the diagram shown in Figure 5.12(b), A is conditionally independent of both B and D given C because the nodes for B and D are nondescendants of node A. The conditional independence assumption made by a na¨ıve Bayes classiﬁer can also be represented using a Bayesian network, as shown in Figure 5.12(c), where y is the target class and {X1 , X2 , . . . , Xd } is the attribute set. Besides the conditional independence conditions imposed by the network topology, each node is also associated with a probability table. 1. If a node X does not have any parents, then the table contains only the prior probability P (X). 2. If a node X has only one parent, Y , then the table contains the conditional probability P (XY ). 3. If a node X has multiple parents, {Y1 , Y2 , . . . , Yk }, then the table contains the conditional probability P (XY1 , Y2 , . . . , Yk ).
241
Chapter 5
Classiﬁcation: Alternative Techniques
E=Yes
D=Healthy
0.7
0.25
Exercise
Diet
HD=Yes E=Yes D=Healthy E=Yes D=Unhealthy E=No D=Healthy E=No D=Unhealthy
HD=Yes HD=No
BP=High 0.85 0.2
D=Healthy D=Unhealthy
0.25 0.45
Heart Disease
0.55
Hb=Yes 0.2 0.85
Heartburn
0.75
CP=Yes
Blood Pressure
Chest Pain
HD=Yes Hb=Yes HD=Yes Hb=No HD=No Hb=Yes HD=No Hb=No
0.8 0.6 0.4 0.1
Figure 5.13. A Bayesian belief network for detecting heart disease and heartburn in patients.
Figure 5.13 shows an example of a Bayesian network for modeling patients with heart disease or heartburn problems. Each variable in the diagram is assumed to be binaryvalued. The parent nodes for heart disease (HD) correspond to risk factors that may aﬀect the disease, such as exercise (E) and diet (D). The child nodes for heart disease correspond to symptoms of the disease, such as chest pain (CP) and high blood pressure (BP). For example, the diagram shows that heartburn (Hb) may result from an unhealthy diet and may lead to chest pain. The nodes associated with the risk factors contain only the prior probabilities, whereas the nodes for heart disease, heartburn, and their corresponding symptoms contain the conditional probabilities. To save space, some of the probabilities have been omitted from the diagram. The omitted probabilities can be recovered by noting that P (X = x) = 1 − P (X = x) and P (X = xY ) = 1 − P (X = xY ), where x denotes the opposite outcome of x. For example, the conditional probability P (Heart Disease = NoExercise = No, Diet = Healthy) = 1 − P (Heart Disease = YesExercise = No, Diet = Healthy) = 1 − 0.55 = 0.45.
242
5.3
Bayesian Classiﬁers
Model Building Model building in Bayesian networks involves two steps: (1) creating the structure of the network, and (2) estimating the probability values in the tables associated with each node. The network topology can be obtained by encoding the subjective knowledge of domain experts. Algorithm 5.3 presents a systematic procedure for inducing the topology of a Bayesian network. Algorithm 5.3 Algorithm for generating the topology of a Bayesian network. 1: Let T = (X1 , X2 , . . . , Xd ) denote a total order of the variables. 2: for j = 1 to d do 3: Let XT (j) denote the j th highest order variable in T . 4: Let π(XT (j) ) = {XT (1) , XT (2) , . . . , XT (j−1) } denote the set of variables preced
ing XT (j) . Remove the variables from π(XT (j) ) that do not aﬀect Xj (using prior knowledge). 6: Create an arc between XT (j) and the remaining variables in π(XT (j) ). 7: end for 5:
Example 5.4. Consider the variables shown in Figure 5.13. After performing Step 1, let us assume that the variables are ordered in the following way: (E, D, HD, Hb, CP, BP ). From Steps 2 to 7, starting with variable D, we obtain the following conditional probabilities: • P (DE) is simpliﬁed to P (D). • P (HDE, D) cannot be simpliﬁed. • P (HbHD, E, D) is simpliﬁed to P (HbD). • P (CP Hb, HD, E, D) is simpliﬁed to P (CP Hb, HD). • P (BP CP, Hb, HD, E, D) is simpliﬁed to P (BP HD). Based on these conditional probabilities, we can create arcs between the nodes (E, HD), (D, HD), (D, Hb), (HD, CP ), (Hb, CP ), and (HD, BP ). These arcs result in the network structure shown in Figure 5.13. Algorithm 5.3 guarantees a topology that does not contain any cycles. The proof for this is quite straightforward. If a cycle exists, then there must be at least one arc connecting the lowerordered nodes to the higherordered nodes, and at least another arc connecting the higherordered nodes to the lowerordered nodes. Since Algorithm 5.3 prevents any arc from connecting the
243
Chapter 5
Classiﬁcation: Alternative Techniques
lowerordered nodes to the higherordered nodes, there cannot be any cycles in the topology. Nevertheless, the network topology may change if we apply a diﬀerent ordering scheme to the variables. Some topology may be inferior because it produces many arcs connecting between diﬀerent pairs of nodes. In principle, we may have to examine all d! possible orderings to determine the most appropriate topology, a task that can be computationally expensive. An alternative approach is to divide the variables into causal and eﬀect variables, and then draw the arcs from each causal variable to its corresponding eﬀect variables. This approach eases the task of building the Bayesian network structure. Once the right topology has been found, the probability table associated with each node is determined. Estimating such probabilities is fairly straightforward and is similar to the approach used by na¨ıve Bayes classiﬁers. Example of Inferencing Using BBN Suppose we are interested in using the BBN shown in Figure 5.13 to diagnose whether a person has heart disease. The following cases illustrate how the diagnosis can be made under diﬀerent scenarios. Case 1: No Prior Information Without any prior information, we can determine whether the person is likely to have heart disease by computing the prior probabilities P (HD = Yes) and P (HD = No). To simplify the notation, let α ∈ {Yes, No} denote the binary values of Exercise and β ∈ {Healthy, Unhealthy} denote the binary values of Diet. P (HD = Yes) =
α
=
α
P (HD = YesE = α, D = β)P (E = α, D = β)
β
P (HD = YesE = α, D = β)P (E = α)P (D = β)
β
= 0.25 × 0.7 × 0.25 + 0.45 × 0.7 × 0.75 + 0.55 × 0.3 × 0.25 + 0.75 × 0.3 × 0.75 = 0.49. Since P (HD = no) = 1 − P (HD = yes) = 0.51, the person has a slightly higher chance of not getting the disease.
244
5.3
Bayesian Classiﬁers
Case 2: High Blood Pressure If the person has high blood pressure, we can make a diagnosis about heart disease by comparing the posterior probabilities, P (HD = YesBP = High) against P (HD = NoBP = High). To do this, we must compute P (BP = High): P (BP = HighHD = γ)P (HD = γ) P (BP = High) = γ
= 0.85 × 0.49 + 0.2 × 0.51 = 0.5185. where γ ∈ {Yes, No}. Therefore, the posterior probability the person has heart disease is P (HD = YesBP = High) = =
P (BP = HighHD = Yes)P (HD = Yes) P (BP = High) 0.85 × 0.49 = 0.8033. 0.5185
Similarly, P (HD = NoBP = High) = 1 − 0.8033 = 0.1967. Therefore, when a person has high blood pressure, it increases the risk of heart disease. Case 3: High Blood Pressure, Healthy Diet, and Regular Exercise Suppose we are told that the person exercises regularly and eats a healthy diet. How does the new information aﬀect our diagnosis? With the new information, the posterior probability that the person has heart disease is P (HD = YesBP = High, D = Healthy, E = Yes) P (BP = HighHD = Yes, D = Healthy, E = Yes) = P (BP = HighD = Healthy, E = Yes) × P (HD = YesD = Healthy, E = Yes) =
P (BP = HighHD = Yes)P (HD = YesD = Healthy, E = Yes) γ P (BP = HighHD = γ)P (HD = γD = Healthy, E = Yes)
=
0.85 × 0.25 0.85 × 0.25 + 0.2 × 0.75
= 0.5862, while the probability that the person does not have heart disease is P (HD = NoBP = High, D = Healthy, E = Yes) = 1 − 0.5862 = 0.4138.
245
Chapter 5
Classiﬁcation: Alternative Techniques
The model therefore suggests that eating healthily and exercising regularly may reduce a person’s risk of getting heart disease. Characteristics of BBN Following are some of the general characteristics of the BBN method: 1. BBN provides an approach for capturing the prior knowledge of a particular domain using a graphical model. The network can also be used to encode causal dependencies among variables. 2. Constructing the network can be time consuming and requires a large amount of eﬀort. However, once the structure of the network has been determined, adding a new variable is quite straightforward. 3. Bayesian networks are well suited to dealing with incomplete data. Instances with missing attributes can be handled by summing or integrating the probabilities over all possible values of the attribute. 4. Because the data is combined probabilistically with prior knowledge, the method is quite robust to model overﬁtting.
5.4
Artiﬁcial Neural Network (ANN)
The study of artiﬁcial neural networks (ANN) was inspired by attempts to simulate biological neural systems. The human brain consists primarily of nerve cells called neurons, linked together with other neurons via strands of ﬁber called axons. Axons are used to transmit nerve impulses from one neuron to another whenever the neurons are stimulated. A neuron is connected to the axons of other neurons via dendrites, which are extensions from the cell body of the neuron. The contact point between a dendrite and an axon is called a synapse. Neurologists have discovered that the human brain learns by changing the strength of the synaptic connection between neurons upon repeated stimulation by the same impulse. Analogous to human brain structure, an ANN is composed of an interconnected assembly of nodes and directed links. In this section, we will examine a family of ANN models, starting with the simplest model called perceptron, and show how the models can be trained to solve classiﬁcation problems.
246
5.4
5.4.1
Artiﬁcial Neural Network (ANN)
Perceptron
Consider the diagram shown in Figure 5.14. The table on the left shows a data set containing three boolean variables (x1 , x2 , x3 ) and an output variable, y, that takes on the value −1 if at least two of the three inputs are zero, and +1 if at least two of the inputs are greater than zero.
X1
X2
X3
y
1 1 1 1 0 0 0 0
0 0 1 1 0 1 1 0
0 1 0 1 1 0 1 0
–1 1 1 1 –1 –1 1 –1
(a) Data set.
Input nodes X1 X2 X3
0.3
Output node
0.3
y
0.3 t = 0.4 (b) Perceptron.
Figure 5.14. Modeling a boolean function using a perceptron.
Figure 5.14(b) illustrates a simple neural network architecture known as a perceptron. The perceptron consists of two types of nodes: input nodes, which are used to represent the input attributes, and an output node, which is used to represent the model output. The nodes in a neural network architecture are commonly known as neurons or units. In a perceptron, each input node is connected via a weighted link to the output node. The weighted link is used to emulate the strength of synaptic connection between neurons. As in biological neural systems, training a perceptron model amounts to adapting the weights of the links until they ﬁt the inputoutput relationships of the underlying data. A perceptron computes its output value, yˆ, by performing a weighted sum on its inputs, subtracting a bias factor t from the sum, and then examining the sign of the result. The model shown in Figure 5.14(b) has three input nodes, each of which has an identical weight of 0.3 to the output node and a bias factor of t = 0.4. The output computed by the model is 1, if 0.3x1 + 0.3x2 + 0.3x3 − 0.4 > 0; yˆ = −1, if 0.3x1 + 0.3x2 + 0.3x3 − 0.4 < 0.
(5.21)
247
Chapter 5
Classiﬁcation: Alternative Techniques
For example, if x1 = 1, x2 = 1, x3 = 0, then yˆ = +1 because 0.3x1 + 0.3x2 + 0.3x3 − 0.4 is positive. On the other hand, if x1 = 0, x2 = 1, x3 = 0, then yˆ = −1 because the weighted sum subtracted by the bias factor is negative. Note the diﬀerence between the input and output nodes of a perceptron. An input node simply transmits the value it receives to the outgoing link without performing any transformation. The output node, on the other hand, is a mathematical device that computes the weighted sum of its inputs, subtracts the bias term, and then produces an output that depends on the sign of the resulting sum. More speciﬁcally, the output of a perceptron model can be expressed mathematically as follows: yˆ = sign wd xd + wd−1 xd−1 + . . . + w2 x2 + w1 x1 − t ,
(5.22)
where w1 , w2 , . . . , wd are the weights of the input links and x1 , x2 , . . . , xd are the input attribute values. The sign function, which acts as an activation function for the output neuron, outputs a value +1 if its argument is positive and −1 if its argument is negative. The perceptron model can be written in a more compact form as follows: yˆ = sign[wd xd + wd−1 xd−1 + . . . + w1 x1 + w0 x0 ] = sign(w · x),
(5.23)
where w0 = −t, x0 = 1, and w · x is the dot product between the weight vector w and the input attribute vector x. Learning Perceptron Model During the training phase of a perceptron model, the weight parameters w are adjusted until the outputs of the perceptron become consistent with the true outputs of training examples. A summary of the perceptron learning algorithm is given in Algorithm 5.4. The key computation for this algorithm is the weight update formula given in Step 7 of the algorithm: (k+1)
wj
(k)
= wj
(k) + λ yi − yˆi xij ,
(5.24)
where w(k) is the weight parameter associated with the ith input link after the k th iteration, λ is a parameter known as the learning rate, and xij is the value of the j th attribute of the training example xi . The justiﬁcation for the weight update formula is rather intuitive. Equation 5.24 shows that the new weight w(k+1) is a combination of the old weight w(k) and a term proportional
248
5.4
Artiﬁcial Neural Network (ANN)
Algorithm 5.4 Perceptron learning algorithm. 1: Let D = {(xi , yi )  i = 1, 2, . . . , N } be the set of training examples. 2: Initialize the weight vector with random values, w(0) 3: repeat 4: for each training example (xi , yi ) ∈ D do 5: 6: 7: 8: 9: 10:
(k)
Compute the predicted output yˆi for each weight wj do (k+1) (k) (k) Update the weight, wj = wj + λ yi − yˆi xij . end for end for until stopping condition is met
to the prediction error, (y − yˆ). If the prediction is correct, then the weight remains unchanged. Otherwise, it is modiﬁed in the following ways: • If y = +1 and yˆ = −1, then the prediction error is (y − yˆ) = 2. To compensate for the error, we need to increase the value of the predicted output by increasing the weights of all links with positive inputs and decreasing the weights of all links with negative inputs. • If yi = −1 and yˆ = +1, then (y − yˆ) = −2. To compensate for the error, we need to decrease the value of the predicted output by decreasing the weights of all links with positive inputs and increasing the weights of all links with negative inputs. In the weight update formula, links that contribute the most to the error term are the ones that require the largest adjustment. However, the weights should not be changed too drastically because the error term is computed only for the current training example. Otherwise, the adjustments made in earlier iterations will be undone. The learning rate λ, a parameter whose value is between 0 and 1, can be used to control the amount of adjustments made in each iteration. If λ is close to 0, then the new weight is mostly inﬂuenced by the value of the old weight. On the other hand, if λ is close to 1, then the new weight is sensitive to the amount of adjustment performed in the current iteration. In some cases, an adaptive λ value can be used; initially, λ is moderately large during the ﬁrst few iterations and then gradually decreases in subsequent iterations. The perceptron model shown in Equation 5.23 is linear in its parameters w and attributes x. Because of this, the decision boundary of a perceptron, which is obtained by setting yˆ = 0, is a linear hyperplane that separates the data into two classes, −1 and +1. Figure 5.15 shows the decision boundary
249
Chapter 5
Classiﬁcation: Alternative Techniques
1 X3 0.5 0 – 0.5
0
0.5 X1
1
0 1.5 – 0.5
0.5 X2
1
1.5
Figure 5.15. Perceptron decision boundary for the data given in Figure 5.14.
obtained by applying the perceptron learning algorithm to the data set given in Figure 5.14. The perceptron learning algorithm is guaranteed to converge to an optimal solution (as long as the learning rate is suﬃciently small) for linearly separable classiﬁcation problems. If the problem is not linearly separable, the algorithm fails to converge. Figure 5.16 shows an example of nonlinearly separable data given by the XOR function. Perceptron cannot ﬁnd the right solution for this data because there is no linear hyperplane that can perfectly separate the training instances. 1.5
1 X1 0 1 0 1
X2
y
0 0 1 1
–1 1 1 –1
X2 0.5
0
– 0.5 – 0.5
0
0.5 X1
1
1.5
Figure 5.16. XOR classification problem. No linear hyperplane can separate the two classes.
250
5.4
5.4.2
Artiﬁcial Neural Network (ANN)
Multilayer Artiﬁcial Neural Network
An artiﬁcial neural network has a more complex structure than that of a perceptron model. The additional complexities may arise in a number of ways: 1. The network may contain several intermediary layers between its input and output layers. Such intermediary layers are called hidden layers and the nodes embedded in these layers are called hidden nodes. The resulting structure is known as a multilayer neural network (see Figure 5.17). In a feedforward neural network, the nodes in one layer X1
X2
X3
X4
X5
Input Layer
Hidden Layer
Output Layer y
Figure 5.17. Example of a multilayer feedforward artificial neural network (ANN).
are connected only to the nodes in the next layer. The perceptron is a singlelayer, feedforward neural network because it has only one layer of nodes—the output layer—that performs complex mathematical operations. In a recurrent neural network, the links may connect nodes within the same layer or nodes from one layer to the previous layers. 2. The network may use types of activation functions other than the sign function. Examples of other activation functions include linear, sigmoid (logistic), and hyperbolic tangent functions, as shown in Figure 5.18. These activation functions allow the hidden and output nodes to produce output values that are nonlinear in their input parameters. These additional complexities allow multilayer neural networks to model more complex relationships between the input and output variables. For ex
251
Chapter 5
Classiﬁcation: Alternative Techniques Linear function
Sigmoid function
1
1
0.5
0.5
0
0
–0.5
– 0.5
1 –1
–0.5
0
0.5
1
1 –1
– 0.5
Tanh function
0
0.5
1
0.5
1
Sign function 1.5
1
1 0.5 0.5 0
0
– 0.5 –0.5 –1 1 –1
–0.5
0
0.5
1
–1.5 –1
– 0.5
0
Figure 5.18. Types of activation functions in artificial neural networks.
ample, consider the XOR problem described in the previous section. The instances can be classiﬁed using two hyperplanes that partition the input space into their respective classes, as shown in Figure 5.19(a). Because a perceptron can create only one hyperplane, it cannot ﬁnd the optimal solution. This problem can be addressed using a twolayer, feedforward neural network, as shown in Figure 5.19(b). Intuitively, we can think of each hidden node as a perceptron that tries to construct one of the two hyperplanes, while the output node simply combines the results of the perceptrons to yield the decision boundary shown in Figure 5.19(a). To learn the weights of an ANN model, we need an eﬃcient algorithm that converges to the right solution when a suﬃcient amount of training data is provided. One approach is to treat each hidden node or output node in the network as an independent perceptron unit and to apply the same weight update formula as Equation 5.24. Obviously, this approach will not work because we lack a priori knowledge about the true outputs of the hidden nodes. This makes it diﬃcult to determine the error term, (y − yˆ), associated
252
5.4
Artiﬁcial Neural Network (ANN) Input Layer
1.5
1
X1
n1
Hidden Layer w31
n3
w41
X2 0.5
Output Layer
w53 n5
0
w32 –0.5 –0.5
0
0.5 X1
1
1.5
X2
(a) Decision boundary.
n2
w42
n4
y
w54
(b) Neural network topology.
Figure 5.19. A twolayer, feedforward neural network for the XOR problem.
with each hidden node. A methodology for learning the weights of a neural network based on the gradient descent approach is presented next. Learning the ANN Model The goal of the ANN learning algorithm is to determine a set of weights w that minimize the total sum of squared errors: 1 (yi − yˆi )2 . 2 N
E(w) =
(5.25)
i=1
Note that the sum of squared errors depends on w because the predicted class yˆ is a function of the weights assigned to the hidden and output nodes. Figure 5.20 shows an example of the error surface as a function of its two parameters, w1 and w2 . This type of error surface is typically encountered when yˆi is a linear function of its parameters, w. If we replace yˆ = w · x into Equation 5.25, then the error function becomes quadratic in its parameters and a global minimum solution can be easily found. In most cases, the output of an ANN is a nonlinear function of its parameters because of the choice of its activation functions (e.g., sigmoid or tanh function). As a result, it is no longer straightforward to derive a solution for w that is guaranteed to be globally optimal. Greedy algorithms such as those based on the gradient descent method have been developed to eﬃciently solve the optimization problem. The weight update formula used by the gradient
253
Chapter 5
Classiﬁcation: Alternative Techniques
E(w1,w2) 1.8 1.6 1.4 1.2 0 1 1 0.5 w1
0 1
0.5 w2
Figure 5.20. Error surface E(w1 , w2 ) for a twoparameter model.
descent method can be written as follows: wj ←− wj − λ
∂E(w) , ∂wj
(5.26)
where λ is the learning rate. The second term states that the weight should be increased in a direction that reduces the overall error term. However, because the error function is nonlinear, it is possible that the gradient descent method may get trapped in a local minimum. The gradient descent method can be used to learn the weights of the output and hidden nodes of a neural network. For hidden nodes, the computation is not trivial because it is diﬃcult to assess their error term, ∂E/∂wj , without knowing what their output values should be. A technique known as backpropagation has been developed to address this problem. There are two phases in each iteration of the algorithm: the forward phase and the backward phase. During the forward phase, the weights obtained from the previous iteration are used to compute the output value of each neuron in the network. The computation progresses in the forward direction; i.e., outputs of the neurons at level k are computed prior to computing the outputs at level k + 1. During the backward phase, the weight update formula is applied in the reverse direction. In other words, the weights at level k + 1 are updated before the weights at level k are updated. This backpropagation approach allows us to use the errors for neurons at layer k + 1 to estimate the errors for neurons at layer k.
254
5.4
Artiﬁcial Neural Network (ANN)
Design Issues in ANN Learning Before we train a neural network to learn a classiﬁcation task, the following design issues must be considered. 1. The number of nodes in the input layer should be determined. Assign an input node to each numerical or binary input variable. If the input variable is categorical, we could either create one node for each categorical value or encode the kary variable using log2 k input nodes. 2. The number of nodes in the output layer should be established. For a twoclass problem, it is suﬃcient to use a single output node. For a kclass problem, there are k output nodes. 3. The network topology (e.g., the number of hidden layers and hidden nodes, and feedforward or recurrent network architecture) must be selected. Note that the target function representation depends on the weights of the links, the number of hidden nodes and hidden layers, biases in the nodes, and type of activation function. Finding the right topology is not an easy task. One way to do this is to start from a fully connected network with a suﬃciently large number of nodes and hidden layers, and then repeat the modelbuilding procedure with a smaller number of nodes. This approach can be very time consuming. Alternatively, instead of repeating the modelbuilding procedure, we could remove some of the nodes and repeat the model evaluation procedure to select the right model complexity. 4. The weights and biases need to be initialized. Random assignments are usually acceptable. 5. Training examples with missing values should be removed or replaced with most likely values.
5.4.3
Characteristics of ANN
Following is a summary of the general characteristics of an artiﬁcial neural network: 1. Multilayer neural networks with at least one hidden layer are universal approximators; i.e., they can be used to approximate any target functions. Since an ANN has a very expressive hypothesis space, it is important to choose the appropriate network topology for a given problem to avoid model overﬁtting.
255
Chapter 5
Classiﬁcation: Alternative Techniques
2. ANN can handle redundant features because the weights are automatically learned during the training step. The weights for redundant features tend to be very small. 3. Neural networks are quite sensitive to the presence of noise in the training data. One approach to handling noise is to use a validation set to determine the generalization error of the model. Another approach is to decrease the weight by some factor at each iteration. 4. The gradient descent method used for learning the weights of an ANN often converges to some local minimum. One way to escape from the local minimum is to add a momentum term to the weight update formula. 5. Training an ANN is a time consuming process, especially when the number of hidden nodes is large. Nevertheless, test examples can be classiﬁed rapidly.
5.5
Support Vector Machine (SVM)
A classiﬁcation technique that has received considerable attention is support vector machine (SVM). This technique has its roots in statistical learning theory and has shown promising empirical results in many practical applications, from handwritten digit recognition to text categorization. SVM also works very well with highdimensional data and avoids the curse of dimensionality problem. Another unique aspect of this approach is that it represents the decision boundary using a subset of the training examples, known as the support vectors. To illustrate the basic idea behind SVM, we ﬁrst introduce the concept of a maximal margin hyperplane and explain the rationale of choosing such a hyperplane. We then describe how a linear SVM can be trained to explicitly look for this type of hyperplane in linearly separable data. We conclude by showing how the SVM methodology can be extended to nonlinearly separable data.
5.5.1
Maximum Margin Hyperplanes
Figure 5.21 shows a plot of a data set containing examples that belong to two diﬀerent classes, represented as squares and circles. The data set is also linearly separable; i.e., we can ﬁnd a hyperplane such that all the squares reside on one side of the hyperplane and all the circles reside on the other
256
5.5
Support Vector Machine (SVM)
Figure 5.21. Possible decision boundaries for a linearly separable data set.
side. However, as shown in Figure 5.21, there are inﬁnitely many such hyperplanes possible. Although their training errors are zero, there is no guarantee that the hyperplanes will perform equally well on previously unseen examples. The classiﬁer must choose one of these hyperplanes to represent its decision boundary, based on how well they are expected to perform on test examples. To get a clearer picture of how the diﬀerent choices of hyperplanes aﬀect the generalization errors, consider the two decision boundaries, B1 and B2 , shown in Figure 5.22. Both decision boundaries can separate the training examples into their respective classes without committing any misclassiﬁcation errors. Each decision boundary Bi is associated with a pair of hyperplanes, denoted as bi1 and bi2 , respectively. bi1 is obtained by moving a parallel hyperplane away from the decision boundary until it touches the closest square(s), whereas bi2 is obtained by moving the hyperplane until it touches the closest circle(s). The distance between these two hyperplanes is known as the margin of the classiﬁer. From the diagram shown in Figure 5.22, notice that the margin for B1 is considerably larger than that for B2 . In this example, B1 turns out to be the maximum margin hyperplane of the training instances. Rationale for Maximum Margin Decision boundaries with large margins tend to have better generalization errors than those with small margins. Intuitively, if the margin is small, then
257
Chapter 5
Classiﬁcation: Alternative Techniques b21 B2 b22
margin for B2
B1 b11 margin for B1
b12
Figure 5.22. Margin of a decision boundary.
any slight perturbations to the decision boundary can have quite a signiﬁcant impact on its classiﬁcation. Classiﬁers that produce decision boundaries with small margins are therefore more susceptible to model overﬁtting and tend to generalize poorly on previously unseen examples. A more formal explanation relating the margin of a linear classiﬁer to its generalization error is given by a statistical learning principle known as structural risk minimization (SRM). This principle provides an upper bound to the generalization error of a classiﬁer (R) in terms of its training error (Re ), the number of training examples (N ), and the model complexity, otherwise known as its capacity (h). More speciﬁcally, with a probability of 1 − η, the generalization error of the classiﬁer can be at worst R ≤ Re + ϕ
h log(η) , , N N
(5.27)
where ϕ is a monotone increasing function of the capacity h. The preceding inequality may seem quite familiar to the readers because it resembles the equation given in Section 4.4.4 (on page 179) for the minimum description length (MDL) principle. In this regard, SRM is another way to express generalization error as a tradeoﬀ between training error and model complexity.
258
5.5
Support Vector Machine (SVM)
The capacity of a linear model is inversely related to its margin. Models with small margins have higher capacities because they are more ﬂexible and can ﬁt many training sets, unlike models with large margins. However, according to the SRM principle, as the capacity increases, the generalization error bound will also increase. Therefore, it is desirable to design linear classiﬁers that maximize the margins of their decision boundaries in order to ensure that their worstcase generalization errors are minimized. One such classiﬁer is the linear SVM, which is explained in the next section.
5.5.2
Linear SVM: Separable Case
A linear SVM is a classiﬁer that searches for a hyperplane with the largest margin, which is why it is often known as a maximal margin classiﬁer. To understand how SVM learns such a boundary, we begin with some preliminary discussion about the decision boundary and margin of a linear classiﬁer. Linear Decision Boundary Consider a binary classiﬁcation problem consisting of N training examples. Each example is denoted by a tuple (xi , yi ) (i = 1, 2, . . . , N ), where xi = (xi1 , xi2 , . . . , xid )T corresponds to the attribute set for the ith example. By convention, let yi ∈ {−1, 1} denote its class label. The decision boundary of a linear classiﬁer can be written in the following form: w · x + b = 0,
(5.28)
where w and b are parameters of the model. Figure 5.23 shows a twodimensional training set consisting of squares and circles. A decision boundary that bisects the training examples into their respective classes is illustrated with a solid line. Any example located along the decision boundary must satisfy Equation 5.28. For example, if xa and xb are two points located on the decision boundary, then w · xa + b = 0, w · xb + b = 0. Subtracting the two equations will yield the following: w · (xb − xa ) = 0,
259
Chapter 5
Classiﬁcation: Alternative Techniques
w.x + b = 0 d x
x1− x2 x1 w
2
w.x + b = 1
w.x + b = −1 O Figure 5.23. Decision boundary and margin of SVM.
where xb − xa is a vector parallel to the decision boundary and is directed from xa to xb . Since the dot product is zero, the direction for w must be perpendicular to the decision boundary, as shown in Figure 5.23. For any square xs located above the decision boundary, we can show that w · xs + b = k,
(5.29)
where k > 0. Similarly, for any circle xc located below the decision boundary, we can show that (5.30) w · xc + b = k , where k < 0. If we label all the squares as class +1 and all the circles as class −1, then we can predict the class label y for any test example z in the following way: 1, if w · z + b > 0; y= (5.31) −1, if w · z + b < 0. Margin of a Linear Classiﬁer Consider the square and the circle that are closest to the decision boundary. Since the square is located above the decision boundary, it must satisfy Equation 5.29 for some positive value k, whereas the circle must satisfy Equation
260
5.5
Support Vector Machine (SVM)
5.30 for some negative value k . We can rescale the parameters w and b of the decision boundary so that the two parallel hyperplanes bi1 and bi2 can be expressed as follows: bi1 : w · x + b = 1,
(5.32)
bi2 : w · x + b = −1.
(5.33)
The margin of the decision boundary is given by the distance between these two hyperplanes. To compute the margin, let x1 be a data point located on bi1 and x2 be a data point on bi2 , as shown in Figure 5.23. Upon substituting these points into Equations 5.32 and 5.33, the margin d can be computed by subtracting the second equation from the ﬁrst equation: w · (x1 − x2 ) = 2 w × d = 2 2 . ∴d= w
(5.34)
Learning a Linear SVM Model The training phase of SVM involves estimating the parameters w and b of the decision boundary from the training data. The parameters must be chosen in such a way that the following two conditions are met: w · xi + b ≥ 1 if yi = 1, w · xi + b ≤ −1 if yi = −1.
(5.35)
These conditions impose the requirements that all training instances from class y = 1 (i.e., the squares) must be located on or above the hyperplane w · x + b = 1, while those instances from class y = −1 (i.e., the circles) must be located on or below the hyperplane w · x + b = −1. Both inequalities can be summarized in a more compact form as follows: yi (w · xi + b) ≥ 1, i = 1, 2, . . . , N.
(5.36)
Although the preceding conditions are also applicable to any linear classiﬁers (including perceptrons), SVM imposes an additional requirement that the margin of its decision boundary must be maximal. Maximizing the margin, however, is equivalent to minimizing the following objective function: f (w) =
w2 . 2
(5.37)
261
Chapter 5
Classiﬁcation: Alternative Techniques
Deﬁnition 5.1 (Linear SVM: Separable Case). The learning task in SVM can be formalized as the following constrained optimization problem: min w
subject to
w2 2
yi (w · xi + b) ≥ 1, i = 1, 2, . . . , N.
Since the objective function is quadratic and the constraints are linear in the parameters w and b, this is known as a convex optimization problem, which can be solved using the standard Lagrange multiplier method. Following is a brief sketch of the main ideas for solving the optimization problem. A more detailed discussion is given in Appendix E. First, we must rewrite the objective function in a form that takes into account the constraints imposed on its solutions. The new objective function is known as the Lagrangian for the optimization problem: N 1 2 λi yi (w · xi + b) − 1 , LP = w − 2
(5.38)
i=1
where the parameters λi are called the Lagrange multipliers. The ﬁrst term in the Lagrangian is the same as the original objective function, while the second term captures the inequality constraints. To understand why the objective function must be modiﬁed, consider the original objective function given in Equation 5.37. It is easy to show that the function is minimized when w = 0, a null vector whose components are all zeros. Such a solution, however, violates the constraints given in Deﬁnition 5.1 because there is no feasible solution for b. The solutions for w and b are infeasible if they violate the inequality constraints; i.e., if yi (w·xi +b)−1 < 0. The Lagrangian given in Equation 5.38 incorporates this constraint by subtracting the term from its original objective function. Assuming that λi ≥ 0, it is clear that any infeasible solution may only increase the value of the Lagrangian. To minimize the Lagrangian, we must take the derivative of LP with respect to w and b and set them to zero: ∂Lp = 0 =⇒ w = λ i yi x i , ∂w
(5.39)
∂Lp = 0 =⇒ λi yi = 0. ∂b
(5.40)
N
i=1
N
i=1
262
5.5
Support Vector Machine (SVM)
Because the Lagrange multipliers are unknown, we still cannot solve for w and b. If Deﬁnition 5.1 contains only equality instead of inequality constraints, then we can use the N equations from equality constraints along with Equations 5.39 and 5.40 to ﬁnd the feasible solutions for w, b, and λi . Note that the Lagrange multipliers for equality constraints are free parameters that can take any values. One way to handle the inequality constraints is to transform them into a set of equality constraints. This is possible as long as the Lagrange multipliers are restricted to be nonnegative. Such transformation leads to the following constraints on the Lagrange multipliers, which are known as the KarushKuhnTucker (KKT) conditions: λi ≥ 0, λi yi (w · xi + b) − 1 = 0.
(5.41) (5.42)
At ﬁrst glance, it may seem that there are as many Lagrange multipliers as there are training instances. It turns out that many of the Lagrange multipliers become zero after applying the constraint given in Equation 5.42. The constraint states that the Lagrange multiplier λi must be zero unless the training instance xi satisﬁes the equation yi (w · xi + b) = 1. Such training instance, with λi > 0, lies along the hyperplanes bi1 or bi2 and is known as a support vector. Training instances that do not reside along these hyperplanes have λi = 0. Equations 5.39 and 5.42 also suggest that the parameters w and b, which deﬁne the decision boundary, depend only on the support vectors. Solving the preceding optimization problem is still quite a daunting task because it involves a large number of parameters: w, b, and λi . The problem can be simpliﬁed by transforming the Lagrangian into a function of the Lagrange multipliers only (this is known as the dual problem). To do this, we ﬁrst substitute Equations 5.39 and 5.40 into Equation 5.38. This will lead to the following dual formulation of the optimization problem: LD =
N i=1
λi −
1 λ i λ j yi yj x i · x j . 2
(5.43)
i,j
The key diﬀerences between the dual and primary Lagrangians are as follows: 1. The dual Lagrangian involves only the Lagrange multipliers and the training data, while the primary Lagrangian involves the Lagrange multipliers as well as parameters of the decision boundary. Nevertheless, the solutions for both optimization problems are equivalent.
263
Chapter 5
Classiﬁcation: Alternative Techniques
2. The quadratic term in Equation 5.43 has a negative sign, which means that the original minimization problem involving the primary Lagrangian, LP , has turned into a maximization problem involving the dual Lagrangian, LD . For large data sets, the dual optimization problem can be solved using numerical techniques such as quadratic programming, a topic that is beyond the scope of this book. Once the λi ’s are found, we can use Equations 5.39 and 5.42 to obtain the feasible solutions for w and b. The decision boundary can be expressed as follows: N
λi yi xi · x + b = 0.
(5.44)
i=1
b is obtained by solving Equation 5.42 for the support vectors. Because the λi ’s are calculated numerically and can have numerical errors, the value computed for b may not be unique. Instead it depends on the support vector used in Equation 5.42. In practice, the average value for b is chosen to be the parameter of the decision boundary. Example 5.5. Consider the twodimensional data set shown in Figure 5.24, which contains eight training instances. Using quadratic programming, we can solve the optimization problem stated in Equation 5.43 to obtain the Lagrange multiplier λi for each training instance. The Lagrange multipliers are depicted in the last column of the table. Notice that only the ﬁrst two instances have nonzero Lagrange multipliers. These instances correspond to the support vectors for this data set. Let w = (w1 , w2 ) and b denote the parameters of the decision boundary. Using Equation 5.39, we can solve for w1 and w2 in the following way: λi yi xi1 = 65.5621 × 1 × 0.3858 + 65.5621 × −1 × 0.4871 = −6.64. w1 = i
w2 =
λi yi xi2 = 65.5621 × 1 × 0.4687 + 65.5621 × −1 × 0.611 = −9.32.
i
The bias term b can be computed using Equation 5.42 for each support vector: b(1) = 1 − w · x1 = 1 − (−6.64)(0.3858) − (−9.32)(0.4687) = 7.9300. b(2) = −1 − w · x2 = −1 − (−6.64)(0.4871) − (−9.32)(0.611) = 7.9289. Averaging these values, we obtain b = 7.93. The decision boundary corresponding to these parameters is shown in Figure 5.24.
264
5.5
Support Vector Machine (SVM)
x1
x2
y
Lagrange Multiplier
0.3858 0.4871 0.9218 0.7382 0.1763 0.4057 0.9355 0.2146
0.4687 0.611 0.4103 0.8936 0.0579 0.3529 0.8132 0.0099
1 –1 –1 –1 1 1 –1 1
65.5261 65.5261 0 0 0 0 0 0
1 0.9
– 6.64 x1 – 9.32 x2 + 7.93 = 0
0.8 0.7
x2
0.6 0.5 0.4 0.3 0.2 0.1 0 0
0.2
0.4
0.6
0.8
1
x1
Figure 5.24. Example of a linearly separable data set.
Once the parameters of the decision boundary are found, a test instance z is classiﬁed as follows: N λ i yi x i · z + b . f (z) = sign w · z + b = sign
i=1
If f (z) = 1, then the test instance is classiﬁed as a positive class; otherwise, it is classiﬁed as a negative class.
265
Chapter 5
5.5.3
Classiﬁcation: Alternative Techniques
Linear SVM: Nonseparable Case
Figure 5.25 shows a data set that is similar to Figure 5.22, except it has two new examples, P and Q. Although the decision boundary B1 misclassiﬁes the new examples, while B2 classiﬁes them correctly, this does not mean that B2 is a better decision boundary than B1 because the new examples may correspond to noise in the training data. B1 should still be preferred over B2 because it has a wider margin, and thus, is less susceptible to overﬁtting. However, the SVM formulation presented in the previous section constructs only decision boundaries that are mistakefree. This section examines how the formulation can be modiﬁed to learn a decision boundary that is tolerable to small training errors using a method known as the soft margin approach. More importantly, the method presented in this section allows SVM to construct a linear decision boundary even in situations where the classes are not linearly separable. To do this, the learning algorithm in SVM must consider the tradeoﬀ between the width of the margin and the number of training errors committed by the linear decision boundary. b21 B2 b22
margin for B2
Q
P B1 b11 margin for B1
b12
Figure 5.25. Decision boundary of SVM for the nonseparable case.
266
5.5
Support Vector Machine (SVM)
w.x + b = 0 1.2 1 0.8
X2
0.6 P
0.4
w.x + b = –1+ξ
ξ /w
0.2 0 w.x + b = –1 –0.2 –0.5
0
0.5 X1
1
1.5
Figure 5.26. Slack variables for nonseparable data.
While the original objective function given in Equation 5.37 is still applicable, the decision boundary B1 no longer satisﬁes all the constraints given in Equation 5.36. The inequality constraints must therefore be relaxed to accommodate the nonlinearly separable data. This can be done by introducing positivevalued slack variables (ξ) into the constraints of the optimization problem, as shown in the following equations: w · xi + b ≥ 1 − ξi if yi = 1, w · xi + b ≤ −1 + ξi if yi = −1,
(5.45)
where ∀i : ξi > 0. To interpret the meaning of the slack variables ξi , consider the diagram shown in Figure 5.26. The circle P is one of the instances that violates the constraints given in Equation 5.35. Let w · x + b = −1 + ξ denote a line that is parallel to the decision boundary and passes through the point P. It can be shown that the distance between this line and the hyperplane w · x + b = −1 is ξ/w. Thus, ξ provides an estimate of the error of the decision boundary on the training example P. In principle, we can apply the same objective function as before and impose the conditions given in Equation 5.45 to ﬁnd the decision boundary. However,
267
Chapter 5
Classiﬁcation: Alternative Techniques
Q
P
Figure 5.27. A decision boundary that has a wide margin but large training error.
since there are no constraints on the number of mistakes the decision boundary can make, the learning algorithm may ﬁnd a decision boundary with a very wide margin but misclassiﬁes many of the training examples, as shown in Figure 5.27. To avoid this problem, the objective function must be modiﬁed to penalize a decision boundary with large values of slack variables. The modiﬁed objective function is given by the following equation: N w2 + C( ξi )k , f (w) = 2 i=1
where C and k are userspeciﬁed parameters representing the penalty of misclassifying the training instances. For the remainder of this section, we assume k = 1 to simplify the problem. The parameter C can be chosen based on the model’s performance on the validation set. It follows that the Lagrangian for this constrained optimization problem can be written as follows: 1 ξi − λi {yi (w · xi + b) − 1 + ξi } − µi ξi , LP = w2 + C 2 N
N
N
i=1
i=1
i=1
(5.46)
where the ﬁrst two terms are the objective function to be minimized, the third term represents the inequality constraints associated with the slack variables,
268
5.5
Support Vector Machine (SVM)
and the last term is the result of the nonnegativity requirements on the values of ξi ’s. Furthermore, the inequality constraints can be transformed into equality constraints using the following KKT conditions: ξi ≥ 0, λi ≥ 0, µi ≥ 0,
(5.47)
λi {yi (w · xi + b) − 1 + ξi } = 0,
(5.48)
µi ξi = 0.
(5.49)
Note that the Lagrange multiplier λi given in Equation 5.48 is nonvanishing only if the training instance resides along the lines w · xi + b = ±1 or has ξi > 0. On the other hand, the Lagrange multipliers µi given in Equation 5.49 are zero for any training instances that are misclassiﬁed (i.e., having ξi > 0). Setting the ﬁrstorder derivative of L with respect to w, b, and ξi to zero would result in the following equations: ∂L = wj − λi yi xij = 0 =⇒ wj = λi yi xij . ∂wj N
N
i=1
∂L =− ∂b
N
(5.50)
i=1
λi yi = 0 =⇒
i=1
N
λi yi = 0.
(5.51)
i=1
∂L = C − λi − µi = 0 =⇒ λi + µi = C. ∂ξi
(5.52)
Substituting Equations 5.50, 5.51, and 5.52 into the Lagrangian will produce the following dual Lagrangian: LD =
1 λ i λ j yi yj x i · x j + C ξi 2 i,j i − λi {yi ( λj yj xi · xj + b) − 1 + ξi } i
j
(C − λi )ξi − i
=
N i=1
λi −
1 λ i λ j yi yj x i · x j , 2
(5.53)
i,j
which turns out to be identical to the dual Lagrangian for linearly separable data (see Equation 5.40 on page 262). Nevertheless, the constraints imposed
269
Chapter 5
Classiﬁcation: Alternative Techniques
on the Lagrange multipliers λi ’s are slightly diﬀerent those in the linearly separable case. In the linearly separable case, the Lagrange multipliers must be nonnegative, i.e., λi ≥ 0. On the other hand, Equation 5.52 suggests that λi should not exceed C (since both µi and λi are nonnegative). Therefore, the Lagrange multipliers for nonlinearly separable data are restricted to 0 ≤ λi ≤ C. The dual problem can then be solved numerically using quadratic programming techniques to obtain the Lagrange multipliers λi . These multipliers can be replaced into Equation 5.50 and the KKT conditions to obtain the parameters of the decision boundary.
5.5.4
Nonlinear SVM
The SVM formulations described in the previous sections construct a linear decision boundary to separate the training examples into their respective classes. This section presents a methodology for applying SVM to data sets that have nonlinear decision boundaries. The trick here is to transform the data from its original coordinate space in x into a new space Φ(x) so that a linear decision boundary can be used to separate the instances in the transformed space. After doing the transformation, we can apply the methodology presented in the previous sections to ﬁnd a linear decision boundary in the transformed space. Attribute Transformation To illustrate how attribute transformation can lead to a linear decision boundary, Figure 5.28(a) shows an example of a twodimensional data set consisting of squares (classiﬁed as y = 1) and circles (classiﬁed as y = −1). The data set is generated in such a way that all the circles are clustered near the center of the diagram and all the squares are distributed farther away from the center. Instances of the data set can be classiﬁed using the following equation: 1 if (x1 − 0.5)2 + (x2 − 0.5)2 > 0.2, (5.54) y(x1 , x2 ) = −1 otherwise. The decision boundary for the data can therefore be written as follows: (x1 − 0.5)2 + (x2 − 0.5)2 = 0.2, which can be further simpliﬁed into the following quadratic equation: x21 − x1 + x22 − x2 = −0.46.
270
5.5
Support Vector Machine (SVM) 0
1 0.9
–0.05
0.8 0.7
–0.1 X2 – X 2
0.5
2
X2
0.6
–0.15
0.4 –0.2
0.3 0.2
–0.25
0.1 0 0
0.1
0.2
0.3
0.4
0.5 X1
0.6
0.7
0.8
0.9
1
(a) Decision boundary in the original twodimensional space.
–0.25
–0.2
–0.15
–0.1
–0.05
0
2
X 1 – X1
(b) Decision boundary in the transformed space.
Figure 5.28. Classifying data with a nonlinear decision boundary.
A nonlinear transformation Φ is needed to map the data from its original feature space into a new space where the decision boundary becomes linear. Suppose we choose the following transformation: √ √ Φ : (x1 , x2 ) −→ (x21 , x22 , 2x1 , 2x2 , 1). (5.55) In the transformed space, we can ﬁnd the parameters w = (w0 , w1 , . . ., w4 ) such that: √ √ w4 x21 + w3 x22 + w2 2x1 + w1 2x2 + w0 = 0. For illustration purposes, let us plot the graph of x22 − x2 versus x21 − x1 for the previously given instances. Figure 5.28(b) shows that in the transformed space, all the circles are located in the lower righthand side of the diagram. A linear decision boundary can therefore be constructed to separate the instances into their respective classes. One potential problem with this approach is that it may suﬀer from the curse of dimensionality problem often associated with highdimensional data. We will show how nonlinear SVM avoids this problem (using a method known as the kernel trick) later in this section. Learning a Nonlinear SVM Model Although the attribute transformation approach seems promising, it raises several implementation issues. First, it is not clear what type of mapping
271
Chapter 5
Classiﬁcation: Alternative Techniques
function should be used to ensure that a linear decision boundary can be constructed in the transformed space. One possibility is to transform the data into an inﬁnite dimensional space, but such a highdimensional space may not be that easy to work with. Second, even if the appropriate mapping function is known, solving the constrained optimization problem in the highdimensional feature space is a computationally expensive task. To illustrate these issues and examine the ways they can be addressed, let us assume that there is a suitable function, Φ(x), to transform a given data set. After the transformation, we need to construct a linear decision boundary that will separate the instances into their respective classes. The linear decision boundary in the transformed space has the following form: w · Φ(x) + b = 0. Deﬁnition 5.2 (Nonlinear SVM). The learning task for a nonlinear SVM can be formalized as the following optimization problem: min w
w2 2
yi (w · Φ(xi ) + b) ≥ 1, i = 1, 2, . . . , N.
subject to
Note the similarity between the learning task of a nonlinear SVM to that of a linear SVM (see Deﬁnition 5.1 on page 262). The main diﬀerence is that, instead of using the original attributes x, the learning task is performed on the transformed attributes Φ(x). Following the approach taken in Sections 5.5.2 and 5.5.3 for linear SVM, we may derive the following dual Lagrangian for the constrained optimization problem: LD =
n
λi −
i=1
1 λi λj yi yj Φ(xi ) · Φ(xj ) 2
(5.56)
i,j
Once the λi ’s are found using quadratic programming techniques, the parameters w and b can be derived using the following equations: w=
λi yi Φ(xi )
(5.57)
i
λj yj Φ(xj ) · Φ(xi ) + b) − 1} = 0, λi {yi ( j
272
(5.58)
5.5
Support Vector Machine (SVM)
which are analogous to Equations 5.39 and 5.40 for linear SVM. Finally, a test instance z can be classiﬁed using the following equation: n λi yi Φ(xi ) · Φ(z) + b . f (z) = sign w · Φ(z) + b = sign
(5.59)
i=1
Except for Equation 5.57, note that the rest of the computations (Equations 5.58 and 5.59) involve calculating the dot product (i.e., similarity) between pairs of vectors in the transformed space, Φ(xi ) · Φ(xj ). Such computation can be quite cumbersome and may suﬀer from the curse of dimensionality problem. A breakthrough solution to this problem comes in the form of a method known as the kernel trick. Kernel Trick The dot product is often regarded as a measure of similarity between two input vectors. For example, the cosine similarity described in Section 2.4.5 on page 73 can be deﬁned as the dot product between two vectors that are normalized to unit length. Analogously, the dot product Φ(xi ) · Φ(xj ) can also be regarded as a measure of similarity between two instances, xi and xj , in the transformed space. The kernel trick is a method for computing similarity in the transformed space using the original attribute set. Consider the mapping function Φ given in Equation 5.55. The dot product between two input vectors u and v in the transformed space can be written as follows: √ √ √ √ Φ(u) · Φ(v) = (u21 , u22 , 2u1 , 2u2 , 1) · (v12 , v22 , 2v1 , 2v2 , 1) = u21 v12 + u22 v22 + 2u1 v1 + 2u2 v2 + 1 = (u · v + 1)2 .
(5.60)
This analysis shows that the dot product in the transformed space can be expressed in terms of a similarity function in the original space: K(u, v) = Φ(u) · Φ(v) = (u · v + 1)2 .
(5.61)
The similarity function, K, which is computed in the original attribute space, is known as the kernel function. The kernel trick helps to address some of the concerns about how to implement nonlinear SVM. First, we do not have to know the exact form of the mapping function Φ because the kernel
273
Chapter 5
Classiﬁcation: Alternative Techniques
functions used in nonlinear SVM must satisfy a mathematical principle known as Mercer’s theorem. This principle ensures that the kernel functions can always be expressed as the dot product between two input vectors in some highdimensional space. The transformed space of the SVM kernels is called a reproducing kernel Hilbert space (RKHS). Second, computing the dot products using kernel functions is considerably cheaper than using the transformed attribute set Φ(x). Third, since the computations are performed in the original space, issues associated with the curse of dimensionality problem can be avoided. Figure 5.29 shows the nonlinear decision boundary obtained by SVM using the polynomial kernel function given in Equation 5.61. A test instance x is classiﬁed according to the following equation: n λi yi Φ(xi ) · Φ(z) + b) f (z) = sign( i=1 n = sign( λi yi K(xi , z) + b) i=1 n = sign( λi yi (xi · z + 1)2 + b),
(5.62)
i=1
where b is the parameter obtained using Equation 5.58. The decision boundary obtained by nonlinear SVM is quite close to the true decision boundary shown in Figure 5.28(a). Mercer’s Theorem The main requirement for the kernel function used in nonlinear SVM is that there must exist a corresponding transformation such that the kernel function computed for a pair of vectors is equivalent to the dot product between the vectors in the transformed space. This requirement can be formally stated in the form of Mercer’s theorem. Theorem 5.1 (Mercer’s Theorem). A kernel function K can be expressed as K(u, v) = Φ(u) · Φ(v) if and only if, for any function g(x) such that g(x)2 dx is ﬁnite, then K(x, y) g(x) g(y) dx dy ≥ 0.
274
5.5
Support Vector Machine (SVM)
1 0.9 0.8 0.7
X2
0.6 0.5 0.4 0.3 0.2 0.1 0 0
0.1
0.2
0.3
0.4
0.5 X1
0.6
0.7
0.8
0.9
1
Figure 5.29. Decision boundary produced by a nonlinear SVM with polynomial kernel.
Kernel functions that satisfy Theorem 5.1 are called positive deﬁnite kernel functions. Examples of such functions are listed below: K(x, y) = (x · y + 1)p −x−y2 /(2σ 2 )
(5.63)
K(x, y) = e
(5.64)
K(x, y) = tanh(kx · y − δ)
(5.65)
Example 5.6. Consider the polynomial kernel function given in Equation 5.63. Let g(x) be a function that has a ﬁnite L2 norm, i.e., g(x)2 dx < ∞. (x · y + 1)p g(x)g(y)dxdy =
p p (x · y)i g(x)g(y)dxdy i i=0
=
p i=0
p i
α1 ,α2 ,...
i α1 α2 α3 (x1 y1 ) (x2 y2 ) (x3 y3 ) . . . α1 α2 . . .
g(x1 , x2 , . . .) g(y1 , y2 , . . .)dx1 dx2 . . . dy1 dy2 . . .
275
Chapter 5
Classiﬁcation: Alternative Techniques
2 p p i α1 α2 = x1 x2 . . . g(x1 , x2 , . . .)dx1 dx2 . . . . i α1 α2 . . . α ,α ,... i=0
1
2
Because the result of the integration is nonnegative, the polynomial kernel function therefore satisﬁes Mercer’s theorem.
5.5.5
Characteristics of SVM
SVM has many desirable qualities that make it one of the most widely used classiﬁcation algorithms. Following is a summary of the general characteristics of SVM: 1. The SVM learning problem can be formulated as a convex optimization problem, in which eﬃcient algorithms are available to ﬁnd the global minimum of the objective function. Other classiﬁcation methods, such as rulebased classiﬁers and artiﬁcial neural networks, employ a greedybased strategy to search the hypothesis space. Such methods tend to ﬁnd only locally optimum solutions. 2. SVM performs capacity control by maximizing the margin of the decision boundary. Nevertheless, the user must still provide other parameters such as the type of kernel function to use and the cost function C for introducing each slack variable. 3. SVM can be applied to categorical data by introducing dummy variables for each categorical attribute value present in the data. For example, if Marital Status has three values {Single, Married, Divorced}, we can introduce a binary variable for each of the attribute values. 4. The SVM formulation presented in this chapter is for binary class problems. Some of the methods available to extend SVM to multiclass problems are presented in Section 5.8.
5.6
Ensemble Methods
The classiﬁcation techniques we have seen so far in this chapter, with the exception of the nearestneighbor method, predict the class labels of unknown examples using a single classiﬁer induced from training data. This section presents techniques for improving classiﬁcation accuracy by aggregating the predictions of multiple classiﬁers. These techniques are known as the ensemble or classiﬁer combination methods. An ensemble method constructs a
276
5.6
Ensemble Methods
set of base classiﬁers from training data and performs classiﬁcation by taking a vote on the predictions made by each base classiﬁer. This section explains why ensemble methods tend to perform better than any single classiﬁer and presents techniques for constructing the classiﬁer ensemble.
5.6.1
Rationale for Ensemble Method
The following example illustrates how an ensemble method can improve a classiﬁer’s performance. Example 5.7. Consider an ensemble of twentyﬁve binary classiﬁers, each of which has an error rate of = 0.35. The ensemble classiﬁer predicts the class label of a test example by taking a majority vote on the predictions made by the base classiﬁers. If the base classiﬁers are identical, then the ensemble will misclassify the same examples predicted incorrectly by the base classiﬁers. Thus, the error rate of the ensemble remains 0.35. On the other hand, if the base classiﬁers are independent—i.e., their errors are uncorrelated—then the ensemble makes a wrong prediction only if more than half of the base classiﬁers predict incorrectly. In this case, the error rate of the ensemble classiﬁer is
eensemble
25 25 i (1 − )25−i = 0.06, = i
(5.66)
i=13
which is considerably lower than the error rate of the base classiﬁers. Figure 5.30 shows the error rate of an ensemble of twentyﬁve binary classiﬁers (eensemble ) for diﬀerent base classiﬁer error rates ( ). The diagonal line represents the case in which the base classiﬁers are identical, while the solid line represents the case in which the base classiﬁers are independent. Observe that the ensemble classiﬁer performs worse than the base classiﬁers when is larger than 0.5. The preceding example illustrates two necessary conditions for an ensemble classiﬁer to perform better than a single classiﬁer: (1) the base classiﬁers should be independent of each other, and (2) the base classiﬁers should do better than a classiﬁer that performs random guessing. In practice, it is diﬃcult to ensure total independence among the base classiﬁers. Nevertheless, improvements in classiﬁcation accuracies have been observed in ensemble methods in which the base classiﬁers are slightly correlated.
277
Chapter 5
Classiﬁcation: Alternative Techniques 1 0.9
Ensemble classifier error
0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0
0.2
0.4 0.6 Base classifier error
0.8
1
Figure 5.30. Comparison between errors of base classifiers and errors of the ensemble classifier.
D
Step 1: Create Multiple Data Sets
D1
D2
Step 2: Build Multiple Classifiers
C1
C2
Step 3: Combine Classifiers
....
Original Training data
Dt1
Dt
Ct1
Ct
C*
Figure 5.31. A logical view of the ensemble learning method.
5.6.2
Methods for Constructing an Ensemble Classiﬁer
A logical view of the ensemble method is presented in Figure 5.31. The basic idea is to construct multiple classiﬁers from the original data and then aggregate their predictions when classifying unknown examples. The ensemble of classiﬁers can be constructed in many ways:
278
5.6
Ensemble Methods
1. By manipulating the training set. In this approach, multiple training sets are created by resampling the original data according to some sampling distribution. The sampling distribution determines how likely it is that an example will be selected for training, and it may vary from one trial to another. A classiﬁer is then built from each training set using a particular learning algorithm. Bagging and boosting are two examples of ensemble methods that manipulate their training sets. These methods are described in further detail in Sections 5.6.4 and 5.6.5. 2. By manipulating the input features. In this approach, a subset of input features is chosen to form each training set. The subset can be either chosen randomly or based on the recommendation of domain experts. Some studies have shown that this approach works very well with data sets that contain highly redundant features. Random forest, which is described in Section 5.6.6, is an ensemble method that manipulates its input features and uses decision trees as its base classiﬁers. 3. By manipulating the class labels. This method can be used when the number of classes is suﬃciently large. The training data is transformed into a binary class problem by randomly partitioning the class labels into two disjoint subsets, A0 and A1 . Training examples whose class label belongs to the subset A0 are assigned to class 0, while those that belong to the subset A1 are assigned to class 1. The relabeled examples are then used to train a base classiﬁer. By repeating the classrelabeling and modelbuilding steps multiple times, an ensemble of base classiﬁers is obtained. When a test example is presented, each base classiﬁer Ci is used to predict its class label. If the test example is predicted as class 0, then all the classes that belong to A0 will receive a vote. Conversely, if it is predicted to be class 1, then all the classes that belong to A1 will receive a vote. The votes are tallied and the class that receives the highest vote is assigned to the test example. An example of this approach is the errorcorrecting output coding method described on page 307. 4. By manipulating the learning algorithm. Many learning algorithms can be manipulated in such a way that applying the algorithm several times on the same training data may result in diﬀerent models. For example, an artiﬁcial neural network can produce diﬀerent models by changing its network topology or the initial weights of the links between neurons. Similarly, an ensemble of decision trees can be constructed by injecting randomness into the treegrowing procedure. For
279
Chapter 5
Classiﬁcation: Alternative Techniques
example, instead of choosing the best splitting attribute at each node, we can randomly choose one of the top k attributes for splitting. The ﬁrst three approaches are generic methods that are applicable to any classiﬁers, whereas the fourth approach depends on the type of classiﬁer used. The base classiﬁers for most of these approaches can be generated sequentially (one after another) or in parallel (all at once). Algorithm 5.5 shows the steps needed to build an ensemble classiﬁer in a sequential manner. The ﬁrst step is to create a training set from the original data D. Depending on the type of ensemble method used, the training sets are either identical to or slight modiﬁcations of D. The size of the training set is often kept the same as the original data, but the distribution of examples may not be identical; i.e., some examples may appear multiple times in the training set, while others may not appear even once. A base classiﬁer Ci is then constructed from each training set Di . Ensemble methods work better with unstable classiﬁers, i.e., base classiﬁers that are sensitive to minor perturbations in the training set. Examples of unstable classiﬁers include decision trees, rulebased classiﬁers, and artiﬁcial neural networks. As will be discussed in Section 5.6.3, the variability among training examples is one of the primary sources of errors in a classiﬁer. By aggregating the base classiﬁers built from diﬀerent training sets, this may help to reduce such types of errors. Finally, a test example x is classiﬁed by combining the predictions made by the base classiﬁers Ci (x): C ∗ (x) = V ote(C1 (x), C2 (x), . . . , Ck (x)). The class can be obtained by taking a majority vote on the individual predictions or by weighting each prediction with the accuracy of the base classiﬁer. Algorithm 5.5 General procedure for ensemble method. 1: Let D denote the original training data, k denote the number of base classiﬁers, 2: 3: 4: 5: 6: 7: 8:
280
and T be the test data. for i = 1 to k do Create training set, Di from D. Build a base classiﬁer Ci from Di . end for for each test record x ∈ T do C ∗ (x) = V ote(C1 (x), C2 (x), . . . , Ck (x)) end for
5.6
5.6.3
Ensemble Methods
BiasVariance Decomposition
Biasvariance decomposition is a formal method for analyzing the prediction error of a predictive model. The following example gives an intuitive explanation for this method. Figure 5.32 shows the trajectories of a projectile launched at a particular angle. Suppose the projectile hits the ﬂoor surface at some location x, at a distance d away from the target position t. Depending on the force applied to the projectile, the observed distance may vary from one trial to another. The observed distance can be decomposed into several components. The ﬁrst component, which is known as bias, measures the average distance between the target position and the location where the projectile hits the ﬂoor. The amount of bias depends on the angle of the projectile launcher. The second component, which is known as variance, measures the deviation between x and the average position x where the projectile hits the ﬂoor. The variance can be explained as a result of changes in the amount of force applied to the projectile. Finally, if the target is not stationary, then the observed distance is also aﬀected by changes in the location of the target. This is considered the noise component associated with variability in the target position. Putting these components together, the average distance can be expressed as: df,θ (y, t) = Biasθ + Variancef + Noiset ,
(5.67)
where f refers to the amount of force applied and θ is the angle of the launcher. The task of predicting the class label of a given example can be analyzed using the same approach. For a given classiﬁer, some predictions may turn out to be correct, while others may be completely oﬀ the mark. We can decompose the expected error of a classiﬁer as a sum of the three terms given in Equation 5.67, where expected error is the probability that the classiﬁer misclassiﬁes a
Target, t y ʻVarianceʼ
ʻNoiseʼ ʻBiasʼ
Figure 5.32. Biasvariance decomposition.
281
Chapter 5
Classiﬁcation: Alternative Techniques
given example. The remainder of this section examines the meaning of bias, variance, and noise in the context of classiﬁcation. A classiﬁer is usually trained to minimize its training error. However, to be useful, the classiﬁer must be able to make an informed guess about the class labels of examples it has never seen before. This requires the classiﬁer to generalize its decision boundary to regions where there are no training examples available—a decision that depends on the design choice of the classiﬁer. For example, a key design issue in decision tree induction is the amount of pruning needed to obtain a tree with low expected error. Figure 5.33 shows two decision trees, T1 and T2 , that are generated from the same training data, but have diﬀerent complexities. T2 is obtained by pruning T1 until a tree with maximum depth of two is obtained. T1 , on the other hand, performs very little pruning on its decision tree. These design choices will introduce a bias into the classiﬁer that is analogous to the bias of the projectile launcher described in the previous example. In general, the stronger the assumptions made by a classiﬁer about the nature of its decision boundary, the larger the classiﬁer’s bias will be. T2 therefore has a larger bias because it makes stronger assumptions about its decision boundary (which is reﬂected by the size of the tree) compared to T1 . Other design choices that may introduce a bias into a classiﬁer include the network topology of an artiﬁcial neural network and the number of neighbors considered by a nearestneighbor classiﬁer. The expected error of a classiﬁer is also aﬀected by variability in the training data because diﬀerent compositions of the training set may lead to diﬀerent decision boundaries. This is analogous to the variance in x when diﬀerent amounts of force are applied to the projectile. The last component of the expected error is associated with the intrinsic noise in the target class. The target class for some domains can be nondeterministic; i.e., instances with the same attribute values can have diﬀerent class labels. Such errors are unavoidable even when the true decision boundary is known. The amount of bias and variance contributing to the expected error depend on the type of classiﬁer used. Figure 5.34 compares the decision boundaries produced by a decision tree and a 1nearest neighbor classiﬁer. For each classiﬁer, we plot the decision boundary obtained by “averaging” the models induced from 100 training sets, each containing 100 examples. The true decision boundary from which the data is generated is also plotted using a dashed line. The diﬀerence between the true decision boundary and the “averaged” decision boundary reﬂects the bias of the classiﬁer. After averaging the models, observe that the diﬀerence between the true decision boundary and the decision boundary produced by the 1nearest neighbor classiﬁer is smaller than
282
5.6 x2 < 1.94
15
x1 < 11.00
x1 < –1.24
Ensemble Methods
10
x2 < 9.25
x2 < 7.45
5 x1 < 1.58 0 –5 –5
(a) Decision tree T1
0
5
10
15
0
5
10
15
15
x2 < 1.94 x1 < –1.24
x1 < 11.00
10 5 0
(b) Decision tree T2
–5 –5
Figure 5.33. Two decision trees with different complexities induced from the same training data.
the observed diﬀerence for a decision tree classiﬁer. This result suggests that the bias of a 1nearest neighbor classiﬁer is lower than the bias of a decision tree classiﬁer. On the other hand, the 1nearest neighbor classiﬁer is more sensitive to the composition of its training examples. If we examine the models induced from diﬀerent training sets, there is more variability in the decision boundary of a 1nearest neighbor classiﬁer than a decision tree classiﬁer. Therefore, the decision boundary of a decision tree classiﬁer has a lower variance than the 1nearest neighbor classiﬁer.
5.6.4
Bagging
Bagging, which is also known as bootstrap aggregating, is a technique that repeatedly samples (with replacement) from a data set according to a uniform probability distribution. Each bootstrap sample has the same size as the original data. Because the sampling is done with replacement, some instances may appear several times in the same training set, while others may be omitted from the training set. On average, a bootstrap sample Di contains approxi
283
Chapter 5
Classiﬁcation: Alternative Techniques
30
30
20
20
10
10
0
0
–10
–10
–20
–20
–30 –30
–20
–10
0
10
20
(a) Decision boundary for decision tree.
30
–30 –30
–20
–10
0
10
20
(b) Decision boundary for 1nearest neighbor.
Figure 5.34. Bias of decision tree and 1nearest neighbor classifiers.
Algorithm 5.6 Bagging algorithm. 1: 2: 3: 4: 5: 6:
Let k be the number of bootstrap samples. for i = 1 to k do Create a bootstrap sample of size N , Di . Train a base classiﬁer Ci on the bootstrap sample Di . end for C ∗ (x) = argmax i δ Ci (x) = y . y
{δ(·) = 1 if its argument is true and 0 otherwise}.
mately 63% of the original training data because each sample has a probability 1 − (1 − 1/N )N of being selected in each Di . If N is suﬃciently large, this probability converges to 1 − 1/e 0.632. The basic procedure for bagging is summarized in Algorithm 5.6. After training the k classiﬁers, a test instance is assigned to the class that receives the highest number of votes. To illustrate how bagging works, consider the data set shown in Table 5.4. Let x denote a onedimensional attribute and y denote the class label. Suppose we apply a classiﬁer that induces only onelevel binary decision trees, with a test condition x ≤ k, where k is a split point chosen to minimize the entropy of the leaf nodes. Such a tree is also known as a decision stump. Without bagging, the best decision stump we can produce splits the records at either x ≤ 0.35 or x ≤ 0.75. Either way, the accuracy of the tree is at
284
30
5.6
Ensemble Methods
Table 5.4. Example of data set used to construct an ensemble of bagging classifiers.
x y
0.1 1
0.2 1
0.3 1
0.4 −1
0.5 −1
0.6 −1
0.7 −1
0.8 1
0.9 1
1 1
most 70%. Suppose we apply the bagging procedure on the data set using ten bootstrap samples. The examples chosen for training in each bagging round are shown in Figure 5.35. On the righthand side of each table, we also illustrate the decision boundary produced by the classiﬁer. We classify the entire data set given in Table 5.4 by taking a majority vote among the predictions made by each base classiﬁer. The results of the predictions are shown in Figure 5.36. Since the class labels are either −1 or +1, taking the majority vote is equivalent to summing up the predicted values of y and examining the sign of the resulting sum (refer to the second to last row in Figure 5.36). Notice that the ensemble classiﬁer perfectly classiﬁes all ten examples in the original data. The preceding example illustrates another advantage of using ensemble methods in terms of enhancing the representation of the target function. Even though each base classiﬁer is a decision stump, combining the classiﬁers can lead to a decision tree of depth 2. Bagging improves generalization error by reducing the variance of the base classiﬁers. The performance of bagging depends on the stability of the base classiﬁer. If a base classiﬁer is unstable, bagging helps to reduce the errors associated with random ﬂuctuations in the training data. If a base classiﬁer is stable, i.e., robust to minor perturbations in the training set, then the error of the ensemble is primarily caused by bias in the base classiﬁer. In this situation, bagging may not be able to improve the performance of the base classiﬁers signiﬁcantly. It may even degrade the classiﬁer’s performance because the eﬀective size of each training set is about 37% smaller than the original data. Finally, since every sample has an equal probability of being selected, bagging does not focus on any particular instance of the training data. It is therefore less susceptible to model overﬁtting when applied to noisy data.
5.6.5
Boosting
Boosting is an iterative procedure used to adaptively change the distribution of training examples so that the base classiﬁers will focus on examples that are hard to classify. Unlike bagging, boosting assigns a weight to each training
285
Chapter 5
Classiﬁcation: Alternative Techniques
Bagging Round 1: x 0.1 0.2 y 1 1
0.2 1
0.3 1
0.4 1
0.4 1
0.5 1
0.6 1
0.9 1
0.9 1
x y = 1 x > 0.35 ==> y = 1
Bagging Round 2: x 0.1 0.2 y 1 1
0.3 1
0.4 1
0.5 1
0.8 1
0.9 1
1 1
1 1
1 1
x y = 1 x > 0.65 ==> y = 1
Bagging Round 3: x 0.1 0.2 y 1 1
0.3 1
0.4 1
0.4 1
0.5 1
0.7 1
0.7 1
0.8 1
0.9 1
x y = 1 x > 0.35 ==> y = 1
Bagging Round 4: x 0.1 0.1 y 1 1
0.2 1
0.4 1
0.4 1
0.5 1
0.5 1
0.7 1
0.8 1
0.9 1
x y = 1 x > 0.3 ==> y = 1
Bagging Round 5: x 0.1 0.1 y 1 1
0.2 1
0.5 1
0.6 1
0.6 1
0.6 1
1 1
1 1
1 1
x y = 1 x > 0.35 ==> y = 1
Bagging Round 6: x 0.2 0.4 y 1 1
0.5 1
0.6 1
0.7 1
0.7 1
0.7 1
0.8 1
0.9 1
1 1
x y = 1 x > 0.75 ==> y = 1
Bagging Round 7: x 0.1 0.4 y 1 1
0.4 1
0.6 1
0.7 1
0.8 1
0.9 1
0.9 1
0.9 1
1 1
x y = 1 x > 0.75 ==> y = 1
Bagging Round 8: x 0.1 0.2 y 1 1
0.5 1
0.5 1
0.5 1
0.7 1
0.7 1
0.8 1
0.9 1
1 1
x y = 1 x > 0.75 ==> y = 1
Bagging Round 9: x 0.1 0.3 y 1 1
0.4 1
0.4 1
0.6 1
0.7 1
0.7 1
0.8 1
1 1
1 1
x y = 1 x > 0.75 ==> y = 1
Bagging Round 10: x 0.1 0.1 y 1 1
0.1 1
0.1 1
0.3 1
0.3 1
0.8 1
0.8 1
0.9 1
0.9 1
x y = 1 x > 0.05 ==> y = 1
Figure 5.35. Example of bagging.
example and may adaptively change the weight at the end of each boosting round. The weights assigned to the training examples can be used in the following ways: 1. They can be used as a sampling distribution to draw a set of bootstrap samples from the original data. 2. They can be used by the base classiﬁer to learn a model that is biased toward higherweight examples.
286
5.6 Round 1 2 3 4 5 6 7 8 9 10 Sum Sign True Class
Ensemble Methods
x=0.1 x=0.2 x=0.3 x=0.4 x=0.5 x=0.6 x=0.7 x=0.8 x=0.9 x=1.0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 2 2 2 6 6 6 6 2 2 2 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
Figure 5.36. Example of combining classifiers constructed using the bagging approach.
This section describes an algorithm that uses weights of examples to determine the sampling distribution of its training set. Initially, the examples are assigned equal weights, 1/N , so that they are equally likely to be chosen for training. A sample is drawn according to the sampling distribution of the training examples to obtain a new training set. Next, a classiﬁer is induced from the training set and used to classify all the examples in the original data. The weights of the training examples are updated at the end of each boosting round. Examples that are classiﬁed incorrectly will have their weights increased, while those that are classiﬁed correctly will have their weights decreased. This forces the classiﬁer to focus on examples that are diﬃcult to classify in subsequent iterations. The following table shows the examples chosen during each boosting round. Boosting (Round 1): Boosting (Round 2): Boosting (Round 3):
7 5 4
3 4 4
2 9 8
8 4 10
7 2 4
9 5 5
4 1 4
10 7 6
6 4 3
3 2 4
Initially, all the examples are assigned the same weights. However, some examples may be chosen more than once, e.g., examples 3 and 7, because the sampling is done with replacement. A classiﬁer built from the data is then used to classify all the examples. Suppose example 4 is diﬃcult to classify. The weight for this example will be increased in future iterations as it gets misclassiﬁed repeatedly. Meanwhile, examples that were not chosen in the pre
287
Chapter 5
Classiﬁcation: Alternative Techniques
vious round, e.g., examples 1 and 5, also have a better chance of being selected in the next round since their predictions in the previous round were likely to be wrong. As the boosting rounds proceed, examples that are the hardest to classify tend to become even more prevalent. The ﬁnal ensemble is obtained by aggregating the base classiﬁers obtained from each boosting round. Over the years, several implementations of the boosting algorithm have been developed. These algorithms diﬀer in terms of (1) how the weights of the training examples are updated at the end of each boosting round, and (2) how the predictions made by each classiﬁer are combined. An implementation called AdaBoost is explored in the next section. AdaBoost Let {(xj , yj )  j = 1, 2, . . . , N } denote a set of N training examples. In the AdaBoost algorithm, the importance of a base classiﬁer Ci depends on its error rate, which is deﬁned as N 1 i = wj I Ci (xj ) = yj , N
(5.68)
j=1
where I(p) = 1 if the predicate p is true, and 0 otherwise. The importance of a classiﬁer Ci is given by the following parameter, 1 αi = ln 2
1 − i . i
Note that αi has a large positive value if the error rate is close to 0 and a large negative value if the error rate is close to 1, as shown in Figure 5.37. The αi parameter is also used to update the weight of the training ex(j) amples. To illustrate, let wi denote the weight assigned to example (xi , yi ) during the j th boosting round. The weight update mechanism for AdaBoost is given by the equation: (j)
(j+1) wi
=
wi × Zj
exp−αj expαj
if Cj (xi ) = yi , if Cj (xi ) = yi
(5.69)
(j+1) where Zj is the normalization factor used to ensure that i wi = 1. The weight update formula given in Equation 5.69 increases the weights of incorrectly classiﬁed examples and decreases the weights of those classiﬁed correctly.
288
5.6
Ensemble Methods
5 4 3
In ((1 – ε)/ε)
2 1 0 –1 –2 –3 –4 –5
0
0.2
0.4
ε
0.6
0.8
1
Figure 5.37. Plot of α as a function of training error .
Instead of using a majority voting scheme, the prediction made by each classiﬁer Cj is weighted according to αj . This approach allows AdaBoost to penalize models that have poor accuracy, e.g., those generated at the earlier boosting rounds. In addition, if any intermediate rounds produce an error rate higher than 50%, the weights are reverted back to their original uniform values, wi = 1/N , and the resampling procedure is repeated. The AdaBoost algorithm is summarized in Algorithm 5.7. Let us examine how the boosting approach works on the data set shown in Table 5.4. Initially, all the examples have identical weights. After three boosting rounds, the examples chosen for training are shown in Figure 5.38(a). The weights for each example are updated at the end of each boosting round using Equation 5.69. Without boosting, the accuracy of the decision stump is, at best, 70%. With AdaBoost, the results of the predictions are given in Figure 5.39(b). The ﬁnal prediction of the ensemble classiﬁer is obtained by taking a weighted average of the predictions made by each base classiﬁer, which is shown in the last row of Figure 5.39(b). Notice that AdaBoost perfectly classiﬁes all the examples in the training data. An important analytical result of boosting shows that the training error of the ensemble is bounded by the following expression: eensemble
≤ i (1 − i ) ,
(5.70)
i
289
Chapter 5
Classiﬁcation: Alternative Techniques
Algorithm 5.7 AdaBoost algorithm. 1: 2: 3: 4: 5: 6: 7: 8: 9: 10: 11: 12: 13: 14: 15:
w = {wj = 1/N  j = 1, 2, . . . , N }. {Initialize the weights for all N examples.} Let k be the number of boosting rounds. for i = 1 to k do Create training set Di by sampling (with replacement) from D according to w. Train a base classiﬁer Ci on Di . Apply C in the i to all examples original training set, D. {Calculate the weighted error.} i = N1 w δ C (x ) = y j i j j j if i > 0.5 then w = {wj = 1/N  j = 1, 2, . . . , N }. {Reset the weights for all N examples.} Go back to Step 4. end if i αi = 12 ln 1− i . Update the weight of each example according to Equation 5.69. end for T C ∗ (x) = argmax j=1 αj δ(Cj (x) = y) . y
where i is the error rate of each base classiﬁer i. If the error rate of the base classiﬁer is less than 50%, we can write i = 0.5 − γi , where γi measures how much better the classiﬁer is than random guessing. The bound on the training error of the ensemble becomes eensemble ≤
γi2 . 1 − 4γi2 ≤ exp − 2 i
(5.71)
i
If γi < γ∗ for all i’s, then the training error of the ensemble decreases exponentially, which leads to the fast convergence of the algorithm. Nevertheless, because of its tendency to focus on training examples that are wrongly classiﬁed, the boosting technique can be quite susceptible to overﬁtting.
5.6.6
Random Forests
Random forest is a class of ensemble methods speciﬁcally designed for decision tree classiﬁers. It combines the predictions made by multiple decision trees, where each tree is generated based on the values of an independent set of random vectors, as shown in Figure 5.40. The random vectors are generated from a ﬁxed probability distribution, unlike the adaptive approach used in AdaBoost, where the probability distribution is varied to focus on examples that are hard to classify. Bagging using decision trees is a special case of random forests, where randomness is injected into the modelbuilding process
290
5.6
Ensemble Methods
Boosting Round 1: x
0.1
0.4
0.5
0.6
0.6
0.7
0.7
0.7
0.8
1
y
1
1
1
1
1
1
1
1
1
1
Boosting Round 2: x
0.1
0.1
0.2
0.2
0.2
0.2
0.3
0.3
0.3
0.3
y
1
1
1
1
1
1
1
1
1
1
Boosting Round 3: x
0.2
0.2
0.4
0.4
0.4
0.4
0.5
0.6
0.6
0.7
y
1
1
1
1
1
1
1
1
1
1
(a) Training records chosen during boosting Round
x=0.1
x=0.2
x=0.3
x=0.4
x=0.5
x=0.6
x=0.7
x=0.8
x=0.9
x=1.0
1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
2
0.311
0.311
0.311
0.01
0.01
0.01
0.01
0.01
0.01
0.01
3
0.029
0.029
0.029
0.228
0.228
0.228
0.228
0.009
0.009
0.009
(b) Weights of training records
Figure 5.38. Example of boosting.
by randomly choosing N samples, with replacement, from the original training set. Bagging also uses the same uniform probability distribution to generate its bootstrapped samples throughout the entire modelbuilding process. It was theoretically proven that the upper bound for generalization error of random forests converges to the following expression, when the number of trees is suﬃciently large. Generalization error ≤
ρ(1 − s2 ) , s2
(5.72)
where ρ is the average correlation among the trees and s is a quantity that measures the “strength” of the tree classiﬁers. The strength of a set of classiﬁers refers to the average performance of the classiﬁers, where performance is measured probabilistically in terms of the classiﬁer’s margin: margin, M (X, Y ) = P (Yˆθ = Y ) − max P (Yˆθ = Z), Z=Y
(5.73)
where Yˆθ is the predicted class of X according to a classiﬁer built from some random vector θ. The higher the margin is, the more likely it is that the
291
Chapter 5
Classiﬁcation: Alternative Techniques
α
Round
Split Point
Left Class
Right Class
1
0.75
1
1
1.738
2
0.05
1
1
2.7784
3
0.3
1
1
4.1195
(a)
Round 1 2 3 Sum Sign
x=0.1 x=0.2 x=0.3 x=0.4 x=0.5 x=0.6 x=0.7 x=0.8 x=0.9 x=1.0 1 1 1 1 1 1 1 1 1 1 1 1
1 1
1 1
1 1
5.16 1
5.16 1
5.16 1
3.08 1
1 1
1 1
3.08 3.08 1 1
1 1
1 1
1 1
1 1
3.08 0.397 0.397 0.397 1 1 1 1
(b)
Figure 5.39. Example of combining classifiers constructed using the AdaBoost approach.
Original Training data
Step 2: Use random vector to build multiple decision trees
Step 3: Combine decision trees
Randomize
D
...
D2
D1
T1
Step 1: Create random vectors
Dt1
T1–1
T2
Dt
T1
T*
Figure 5.40. Random forests.
classiﬁer correctly predicts a given example X. Equation 5.72 is quite intuitive; as the trees become more correlated or the strength of the ensemble decreases, the generalization error bound tends to increase. Randomization helps to reduce the correlation among decision trees so that the generalization error of the ensemble can be improved.
292
5.6
Ensemble Methods
Each decision tree uses a random vector that is generated from some ﬁxed probability distribution. A random vector can be incorporated into the treegrowing process in many ways. The ﬁrst approach is to randomly select F input features to split at each node of the decision tree. As a result, instead of examining all the available features, the decision to split a node is determined from these selected F features. The tree is then grown to its entirety without any pruning. This may help reduce the bias present in the resulting tree. Once the trees have been constructed, the predictions are combined using a majority voting scheme. This approach is known as ForestRI, where RI refers to random input selection. To increase randomness, bagging can also be used to generate bootstrap samples for ForestRI. The strength and correlation of random forests may depend on the size of F . If F is suﬃciently small, then the trees tend to become less correlated. On the other hand, the strength of the tree classiﬁer tends to improve with a larger number of features, F . As a tradeoﬀ, the number of features is commonly chosen to be F = log2 d + 1, where d is the number of input features. Since only a subset of the features needs to be examined at each node, this approach helps to signiﬁcantly reduce the runtime of the algorithm. If the number of original features d is too small, then it is diﬃcult to choose an independent set of random features for building the decision trees. One way to increase the feature space is to create linear combinations of the input features. Speciﬁcally, at each node, a new feature is generated by randomly selecting L of the input features. The input features are linearly combined using coeﬃcients generated from a uniform distribution in the range of [−1, 1]. At each node, F of such randomly combined new features are generated, and the best of them is subsequently selected to split the node. This approach is known as ForestRC. A third approach for generating the random trees is to randomly select one of the F best splits at each node of the decision tree. This approach may potentially generate trees that are more correlated than ForestRI and ForestRC, unless F is suﬃciently large. It also does not have the runtime savings of ForestRI and ForestRC because the algorithm must examine all the splitting features at each node of the decision tree. It has been shown empirically that the classiﬁcation accuracies of random forests are quite comparable to the AdaBoost algorithm. It is also more robust to noise and runs much faster than the AdaBoost algorithm. The classiﬁcation accuracies of various ensemble algorithms are compared in the next section.
293
Chapter 5
Classiﬁcation: Alternative Techniques
Table 5.5. Comparing the accuracy of a decision tree classifier against three ensemble methods. Data Set
Anneal Australia Auto Breast Cleve Credit Diabetes German Glass Heart Hepatitis Horse Ionosphere Iris Labor Led7 Lymphography Pima Sonar Tictactoe Vehicle Waveform Wine Zoo
5.6.7
Number of (Attributes, Classes, Records) (39, 6, 898) (15, 2, 690) (26, 7, 205) (11, 2, 699) (14, 2, 303) (16, 2, 690) (9, 2, 768) (21, 2, 1000) (10, 7, 214) (14, 2, 270) (20, 2, 155) (23, 2, 368) (35, 2, 351) (5, 3, 150) (17, 2, 57) (8, 10, 3200) (19, 4, 148) (9, 2, 768) (61, 2, 208) (10, 2, 958) (19, 4, 846) (22, 3, 5000) (14, 3, 178) (17, 7, 101)
Decision Tree (%)
Bagging (%)
Boosting (%)
RF (%)
92.09 85.51 81.95 95.14 76.24 85.8 72.40 70.90 67.29 80.00 81.94 85.33 89.17 94.67 78.95 73.34 77.03 74.35 78.85 83.72 71.04 76.44 94.38 93.07
94.43 87.10 85.37 96.42 81.52 86.23 76.30 73.40 76.17 81.48 81.29 85.87 92.02 94.67 84.21 73.66 79.05 76.69 78.85 93.84 74.11 83.30 96.07 93.07
95.43 85.22 85.37 97.28 82.18 86.09 73.18 73.00 77.57 80.74 83.87 81.25 93.73 94.00 89.47 73.34 85.14 73.44 84.62 98.54 78.25 83.90 97.75 95.05
95.43 85.80 84.39 96.14 82.18 85.8 75.13 74.5 78.04 83.33 83.23 85.33 93.45 93.33 84.21 73.06 82.43 77.60 85.58 95.82 74.94 84.04 97.75 97.03
Empirical Comparison among Ensemble Methods
Table 5.5 shows the empirical results obtained when comparing the performance of a decision tree classiﬁer against bagging, boosting, and random forest. The base classiﬁers used in each ensemble method consist of ﬁfty decision trees. The classiﬁcation accuracies reported in this table are obtained from tenfold crossvalidation. Notice that the ensemble classiﬁers generally outperform a single decision tree classiﬁer on many of the data sets.
5.7
Class Imbalance Problem
Data sets with imbalanced class distributions are quite common in many real applications. For example, an automated inspection system that monitors products that come oﬀ a manufacturing assembly line may ﬁnd that the num
294
5.7
Class Imbalance Problem
ber of defective products is signiﬁcantly fewer than that of nondefective products. Similarly, in credit card fraud detection, fraudulent transactions are outnumbered by legitimate transactions. In both of these examples, there is a disproportionate number of instances that belong to diﬀerent classes. The degree of imbalance varies from one application to another—a manufacturing plant operating under the six sigma principle may discover four defects in a million products shipped to their customers, while the amount of credit card fraud may be of the order of 1 in 100. Despite their infrequent occurrences, a correct classiﬁcation of the rare class in these applications often has greater value than a correct classiﬁcation of the majority class. However, because the class distribution is imbalanced, this presents a number of problems to existing classiﬁcation algorithms. The accuracy measure, which is used extensively to compare the performance of classiﬁers, may not be well suited for evaluating models derived from imbalanced data sets. For example, if 1% of the credit card transactions are fraudulent, then a model that predicts every transaction as legitimate has an accuracy of 99% even though it fails to detect any of the fraudulent activities. Additionally, measures that are used to guide the learning algorithm (e.g., information gain for decision tree induction) may need to be modiﬁed to focus on the rare class. Detecting instances of the rare class is akin to ﬁnding a needle in a haystack. Because their instances occur infrequently, models that describe the rare class tend to be highly specialized. For example, in a rulebased classiﬁer, the rules extracted for the rare class typically involve a large number of attributes and cannot be easily simpliﬁed into more general rules with broader coverage (unlike the rules for the majority class). Such models are also susceptible to the presence of noise in training data. As a result, many of the existing classiﬁcation algorithms may not eﬀectively detect instances of the rare class. This section presents some of the methods developed for handling the class imbalance problem. First, alternative metrics besides accuracy are introduced, along with a graphical method called ROC analysis. We then describe how costsensitive learning and samplingbased methods may be used to improve the detection of rare classes.
5.7.1
Alternative Metrics
Since the accuracy measure treats every class as equally important, it may not be suitable for analyzing imbalanced data sets, where the rare class is considered more interesting than the majority class. For binary classiﬁcation, the rare class is often denoted as the positive class, while the majority class is
295
Chapter 5
Classiﬁcation: Alternative Techniques
Table 5.6. A confusion matrix for a binary classification problem in which the classes are not equally important. Predicted Class +
−
Actual
+
f++ (TP)
f+− (FN)
Class
−
f−+ (FP)
f−− (TN)
denoted as the negative class. A confusion matrix that summarizes the number of instances predicted correctly or incorrectly by a classiﬁcation model is shown in Table 5.6. The following terminology is often used when referring to the counts tabulated in a confusion matrix: • True positive (TP) or f++ , which corresponds to the number of positive examples correctly predicted by the classiﬁcation model. • False negative (FN) or f+− , which corresponds to the number of positive examples wrongly predicted as negative by the classiﬁcation model. • False positive (FP) or f−+ , which corresponds to the number of negative examples wrongly predicted as positive by the classiﬁcation model. • True negative (TN) or f−− , which corresponds to the number of negative examples correctly predicted by the classiﬁcation model. The counts in a confusion matrix can also be expressed in terms of percentages. The true positive rate (T P R) or sensitivity is deﬁned as the fraction of positive examples predicted correctly by the model, i.e., T P R = T P/(T P + F N ). Similarly, the true negative rate (T N R) or speciﬁcity is deﬁned as the fraction of negative examples predicted correctly by the model, i.e., T N R = T N/(T N + F P ). Finally, the false positive rate (F P R) is the fraction of negative examples predicted as a positive class, i.e., F P R = F P/(T N + F P ),
296
5.7
Class Imbalance Problem
while the false negative rate (F N R) is the fraction of positive examples predicted as a negative class, i.e., F N R = F N/(T P + F N ). Recall and precision are two widely used metrics employed in applications where successful detection of one of the classes is considered more significant than detection of the other classes. A formal deﬁnition of these metrics is given below. TP TP + FP TP Recall, r = TP + FN Precision, p =
(5.74) (5.75)
Precision determines the fraction of records that actually turns out to be positive in the group the classiﬁer has declared as a positive class. The higher the precision is, the lower the number of false positive errors committed by the classiﬁer. Recall measures the fraction of positive examples correctly predicted by the classiﬁer. Classiﬁers with large recall have very few positive examples misclassiﬁed as the negative class. In fact, the value of recall is equivalent to the true positive rate. It is often possible to construct baseline models that maximize one metric but not the other. For example, a model that declares every record to be the positive class will have a perfect recall, but very poor precision. Conversely, a model that assigns a positive class to every test record that matches one of the positive records in the training set has very high precision, but low recall. Building a model that maximizes both precision and recall is the key challenge of classiﬁcation algorithms. Precision and recall can be summarized into another metric known as the F1 measure. 2 × TP 2rp F1 = = (5.76) r+p 2 × TP + FP + FN In principle, F1 represents a harmonic mean between recall and precision, i.e., F1 =
1 r
2 . + p1
The harmonic mean of two numbers x and y tends to be closer to the smaller of the two numbers. Hence, a high value of F1 measure ensures that both
297
Chapter 5
Classiﬁcation: Alternative Techniques
precision and recall are reasonably high. A comparison among harmonic, geometric, and arithmetic means is given in the next example. Example 5.8. Consider two positive numbers a = 1 and b = 5. Their√arithmetic mean is µa = (a + b)/2 = 3 and their geometric mean is µg = ab = 2.236. Their harmonic mean is µh = (2×1×5)/6 = 1.667, which is closer to the smaller value between a and b than the arithmetic and geometric means. More generally, the Fβ measure can be used to examine the tradeoﬀ between recall and precision: Fβ =
(β 2 + 1)rp (β 2 + 1) × T P = . r + β2p (β 2 + 1)T P + β 2 F P + F N
(5.77)
Both precision and recall are special cases of Fβ by setting β = 0 and β = ∞, respectively. Low values of β make Fβ closer to precision, and high values make it closer to recall. A more general metric that captures Fβ as well as accuracy is the weighted accuracy measure, which is deﬁned by the following equation: Weighted accuracy =
w1 T P + w4 T N . w 1 T P + w 2 F P + w 3 F N + w4 T N
(5.78)
The relationship between weighted accuracy and other performance metrics is summarized in the following table: Measure Recall Precision Fβ Accuracy
5.7.2
w1 1 1 β2 + 1 1
w2 1 0 β2 1
w3 0 1 1 1
w4 0 0 0 1
The Receiver Operating Characteristic Curve
A receiver operating characteristic (ROC) curve is a graphical approach for displaying the tradeoﬀ between true positive rate and false positive rate of a classiﬁer. In an ROC curve, the true positive rate (T P R) is plotted along the y axis and the false positive rate (F P R) is shown on the x axis. Each point along the curve corresponds to one of the models induced by the classiﬁer. Figure 5.41 shows the ROC curves for a pair of classiﬁers, M1 and M2 .
298
5.7
Class Imbalance Problem
1 0.9
M2
0.8
True Positive Rate
0.7 0.6
M1
0.5 0.4 0.3 0.2 0.1 0
0.1
0.2
0.3
0.4 0.5 0.6 0.7 False Positive Rate
0.8
0.9
1
Figure 5.41. ROC curves for two different classifiers.
There are several critical points along an ROC curve that have wellknown interpretations: (TPR=0, FPR=0): Model predicts every instance to be a negative class. (TPR=1, FPR=1): Model predicts every instance to be a positive class. (TPR=1, FPR=0): The ideal model. A good classiﬁcation model should be located as close as possible to the upper left corner of the diagram, while a model that makes random guesses should reside along the main diagonal, connecting the points (T P R = 0, F P R = 0) and (T P R = 1, F P R = 1). Random guessing means that a record is classiﬁed as a positive class with a ﬁxed probability p, irrespective of its attribute set. For example, consider a data set that contains n+ positive instances and n− negative instances. The random classiﬁer is expected to correctly classify pn+ of the positive instances and to misclassify pn− of the negative instances. Therefore, the T P R of the classiﬁer is (pn+ )/n+ = p, while its F P R is (pn− )/p = p. Since the T P R and F P R are identical, the ROC curve for a random classiﬁer always reside along the main diagonal. An ROC curve is useful for comparing the relative performance among diﬀerent classiﬁers. In Figure 5.41, M1 is better than M2 when F P R is less
299
Chapter 5
Classiﬁcation: Alternative Techniques
than 0.36, while M2 is superior when F P R is greater than 0.36. Clearly, neither of these two classiﬁers dominates the other. The area under the ROC curve (AUC) provides another approach for evaluating which model is better on average. If the model is perfect, then its area under the ROC curve would equal 1. If the model simply performs random guessing, then its area under the ROC curve would equal 0.5. A model that is strictly better than another would have a larger area under the ROC curve. Generating an ROC curve To draw an ROC curve, the classiﬁer should be able to produce a continuousvalued output that can be used to rank its predictions, from the most likely record to be classiﬁed as a positive class to the least likely record. These outputs may correspond to the posterior probabilities generated by a Bayesian classiﬁer or the numericvalued outputs produced by an artiﬁcial neural network. The following procedure can then be used to generate an ROC curve: 1. Assuming that the continuousvalued outputs are deﬁned for the positive class, sort the test records in increasing order of their output values. 2. Select the lowest ranked test record (i.e., the record with lowest output value). Assign the selected record and those ranked above it to the positive class. This approach is equivalent to classifying all the test records as positive class. Because all the positive examples are classiﬁed correctly and the negative examples are misclassiﬁed, T P R = F P R = 1. 3. Select the next test record from the sorted list. Classify the selected record and those ranked above it as positive, while those ranked below it as negative. Update the counts of T P and F P by examining the actual class label of the previously selected record. If the previously selected record is a positive class, the T P count is decremented and the F P count remains the same as before. If the previously selected record is a negative class, the F P count is decremented and T P count remains the same as before. 4. Repeat Step 3 and update the T P and F P counts accordingly until the highest ranked test record is selected. 5. Plot the T P R against F P R of the classiﬁer. Figure 5.42 shows an example of how to compute the ROC curve. There are ﬁve positive examples and ﬁve negative examples in the test set. The class
300
5.7 Class
Class Imbalance Problem
+
–
+
–
–
–
+
–
+
+
0.25
0.43
0.53
0.76
0.85
0.85
0.85
0.87
0.93
0.95
1.00
TP
5
4
4
3
3
3
3
2
2
1
0
FP
5
5
4
4
3
2
1
1
0
0
0
TN
0
0
1
1
2
3
4
4
5
5
5
FN
0
1
1
2
2
2
2
3
3
4
5
TPR
1
0.8
0.8
0.6
0.6
0.6
0.6
0.4
0.4
0.2
0
FPR
1
1
0.8
0.8
0.6
0.4
0.2
0.2
0
0
0
Figure 5.42. Constructing an ROC curve. 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Figure 5.43. ROC curve for the data shown in Figure 5.42.
labels of the test records are shown in the ﬁrst row of the table. The second row corresponds to the sorted output values for each record. For example, they may correspond to the posterior probabilities P (+x) generated by a na¨ıve Bayes classiﬁer. The next six rows contain the counts of T P , F P , T N , and F N , along with their corresponding T P R and F P R. The table is then ﬁlled from left to right. Initially, all the records are predicted to be positive. Thus, T P = F P = 5 and T P R = F P R = 1. Next, we assign the test record with the lowest output value as the negative class. Because the selected record is actually a positive example, the T P count reduces from 5 to 4 and the F P count is the same as before. The F P R and T P R are updated accordingly. This process is repeated until we reach the end of the list, where T P R = 0 and F P R = 0. The ROC curve for this example is shown in Figure 5.43.
301
Chapter 5
5.7.3
Classiﬁcation: Alternative Techniques
CostSensitive Learning
A cost matrix encodes the penalty of classifying records from one class as another. Let C(i, j) denote the cost of predicting a record from class i as class j. With this notation, C(+, −) is the cost of committing a false negative error, while C(−, +) is the cost of generating a false alarm. A negative entry in the cost matrix represents the reward for making correct classiﬁcation. Given a collection of N test records, the overall cost of a model M is Ct (M ) =
T P × C(+, +) + F P × C(−, +) + F N × C(+, −) + T N × C(−, −).
(5.79)
Under the 0/1 cost matrix, i.e., C(+, +) = C(−, −) = 0 and C(+, −) = C(−, +) = 1, it can be shown that the overall cost is equivalent to the number of misclassiﬁcation errors. Ct (M ) = 0 × (T P + T N ) + 1 × (F P + F N ) = N × Err,
(5.80)
where Err is the error rate of the classiﬁer. Example 5.9. Consider the cost matrix shown in Table 5.7: The cost of committing a false negative error is a hundred times larger than the cost of committing a false alarm. In other words, failure to detect any positive example is just as bad as committing a hundred false alarms. Given the classiﬁcation models with the confusion matrices shown in Table 5.8, the total cost for each model is Ct (M1 ) = 150 × (−1) + 60 × 1 + 40 × 100 = 3910, Ct (M2 ) = 250 × (−1) + 5 × 1 + 45 × 100 = 4255.
Table 5.7. Cost matrix for Example 5.9.
Actual Class
302
Class = + Class = −
Predicted Class Class = + Class = − −1 100 1 0
5.7
Class Imbalance Problem
Table 5.8. Confusion matrix for two classification models. Model M1 Predicted Class Model M2 Predicted Class Class + Class Class + Class Actual Class + 150 40 Actual Class + 250 45 Class Class 60 250 Class Class 5 200
Notice that despite improving both of its true positive and false positive counts, model M2 is still inferior since the improvement comes at the expense of increasing the more costly false negative errors. A standard accuracy measure would have preferred model M2 over M1 . A costsensitive classiﬁcation technique takes the cost matrix into consideration during model building and generates a model that has the lowest cost. For example, if false negative errors are the most costly, the learning algorithm will try to reduce these errors by extending its decision boundary toward the negative class, as shown in Figure 5.44. In this way, the generated model can cover more positive examples, although at the expense of generating additional false alarms.
B2
B1
Figure 5.44. Modifying the decision boundary (from B1 to B2 ) to reduce the false negative errors of a classifier.
There are various ways to incorporate cost information into classiﬁcation algorithms. For example, in the context of decision tree induction, the cost
303
Chapter 5
Classiﬁcation: Alternative Techniques
information can be used to: (1) choose the best attribute to use for splitting the data, (2) determine whether a subtree should be pruned, (3) manipulate the weights of the training records so that the learning algorithm converges to a decision tree that has the lowest cost, and (4) modify the decision rule at each leaf node. To illustrate the last approach, let p(it) denote the fraction of training records from class i that belong to the leaf node t. A typical decision rule for a binary classiﬁcation problem assigns the positive class to node t if the following condition holds. p(+t) > p(−t) =⇒
p(+t) > (1 − p(+t))
=⇒
2p(+t) > 1
=⇒
p(+t) > 0.5.
(5.81)
The preceding decision rule suggests that the class label of a leaf node depends on the majority class of the training records that reach the particular node. Note that this rule assumes that the misclassiﬁcation costs are identical for both positive and negative examples. This decision rule is equivalent to the expression given in Equation 4.8 on page 165. Instead of taking a majority vote, a costsensitive algorithm assigns the class label i to node t if it minimizes the following expression: C(it) =
p(jt)C(j, i).
(5.82)
j
In the case where C(+, +) = C(−, −) = 0, a leaf node t is assigned to the positive class if: p(+t)C(+, −) > p(−t)C(−, +) =⇒ =⇒
p(+t)C(+, −) > (1 − p(+t))C(−, +) C(−, +) . p(+t) > C(−, +) + C(+, −)
(5.83)
This expression suggests that we can modify the threshold of the decision rule from 0.5 to C(−, +)/(C(−, +) + C(+, −)) to obtain a costsensitive classiﬁer. If C(−, +) < C(+, −), then the threshold will be less than 0.5. This result makes sense because the cost of making a false negative error is more expensive than that for generating a false alarm. Lowering the threshold will expand the decision boundary toward the negative class, as shown in Figure 5.44.
304
Class Imbalance Problem
x2
x
2
5.7
x
x
(a) Without oversampling
(b) With oversampling
1
1
Figure 5.45. Illustrating the effect of oversampling of the rare class.
5.7.4
SamplingBased Approaches
Sampling is another widely used approach for handling the class imbalance problem. The idea of sampling is to modify the distribution of instances so that the rare class is well represented in the training set. Some of the available techniques for sampling include undersampling, oversampling, and a hybrid of both approaches. To illustrate these techniques, consider a data set that contains 100 positive examples and 1000 negative examples. In the case of undersampling, a random sample of 100 negative examples is chosen to form the training set along with all the positive examples. One potential problem with this approach is that some of the useful negative examples may not be chosen for training, therefore, resulting in a less than optimal model. A potential method to overcome this problem is to perform undersampling multiple times and to induce multiple classiﬁers similar to the ensemble learning approach. Focused undersampling methods may also be used, where the sampling procedure makes an informed choice with regard to the negative examples that should be eliminated, e.g., those located far away from the decision boundary. Oversampling replicates the positive examples until the training set has an equal number of positive and negative examples. Figure 5.45 illustrates the eﬀect of oversampling on the construction of a decision boundary using a classiﬁer such as a decision tree. Without oversampling, only the positive examples at the bottom righthand side of Figure 5.45(a) are classiﬁed correctly. The positive example in the middle of the diagram is misclassiﬁed because there
305
Chapter 5
Classiﬁcation: Alternative Techniques
are not enough examples to justify the creation of a new decision boundary to separate the positive and negative instances. Oversampling provides the additional examples needed to ensure that the decision boundary surrounding the positive example is not pruned, as illustrated in Figure 5.45(b). However, for noisy data, oversampling may cause model overﬁtting because some of the noise examples may be replicated many times. In principle, oversampling does not add any new information into the training set. Replication of positive examples only prevents the learning algorithm from pruning certain parts of the model that describe regions that contain very few training examples (i.e., the small disjuncts). The additional positive examples also tend to increase the computation time for model building. The hybrid approach uses a combination of undersampling the majority class and oversampling the rare class to achieve uniform class distribution. Undersampling can be performed using random or focused subsampling. Oversampling, on the other hand, can be done by replicating the existing positive examples or generating new positive examples in the neighborhood of the existing positive examples. In the latter approach, we must ﬁrst determine the knearest neighbors for each existing positive example. A new positive example is then generated at some random point along the line segment that joins the positive example to one of its knearest neighbors. This process is repeated until the desired number of positive examples is reached. Unlike the data replication approach, the new examples allow us to extend the decision boundary for the positive class outward, similar to the approach shown in Figure 5.44. Nevertheless, this approach may still be quite susceptible to model overﬁtting.
5.8
Multiclass Problem
Some of the classiﬁcation techniques described in this chapter, such as support vector machines and AdaBoost, are originally designed for binary classiﬁcation problems. Yet there are many realworld problems, such as character recognition, face identiﬁcation, and text classiﬁcation, where the input data is divided into more than two categories. This section presents several approaches for extending the binary classiﬁers to handle multiclass problems. To illustrate these approaches, let Y = {y1 , y2 , . . . , yK } be the set of classes of the input data. The ﬁrst approach decomposes the multiclass problem into K binary problems. For each class yi ∈ Y , a binary problem is created where all instances that belong to yi are considered positive examples, while the remaining in
306
5.8
Multiclass Problem
stances are considered negative examples. A binary classiﬁer is then constructed to separate instances of class yi from the rest of the classes. This is known as the oneagainstrest (1r) approach. The second approach, which is known as the oneagainstone (11) approach, constructs K(K − 1)/2 binary classiﬁers, where each classiﬁer is used to distinguish between a pair of classes, (yi , yj ). Instances that do not belong to either yi or yj are ignored when constructing the binary classiﬁer for (yi , yj ). In both 1r and 11 approaches, a test instance is classiﬁed by combining the predictions made by the binary classiﬁers. A voting scheme is typically employed to combine the predictions, where the class that receives the highest number of votes is assigned to the test instance. In the 1r approach, if an instance is classiﬁed as negative, then all classes except for the positive class receive a vote. This approach, however, may lead to ties among the diﬀerent classes. Another possibility is to transform the outputs of the binary classiﬁers into probability estimates and then assign the test instance to the class that has the highest probability. Example 5.10. Consider a multiclass problem where Y = {y1 , y2 , y3 , y4 }. Suppose a test instance is classiﬁed as (+, −, −, −) according to the 1r approach. In other words, it is classiﬁed as positive when y1 is used as the positive class and negative when y2 , y3 , and y4 are used as the positive class. Using a simple majority vote, notice that y1 receives the highest number of votes, which is four, while the remaining classes receive only three votes. The test instance is therefore classiﬁed as y1 . Suppose the test instance is classiﬁed as follows using the 11 approach: Binary pair of classes Classiﬁcation
+: y1 −: y2 +
+: y1 −: y3 +
+: y1 −: y4 −
+: y2 −: y3 +
+: y2 −: y4 −
+: y3 −: y4 +
The ﬁrst two rows in this table correspond to the pair of classes (yi , yj ) chosen to build the classiﬁer and the last row represents the predicted class for the test instance. After combining the predictions, y1 and y4 each receive two votes, while y2 and y3 each receives only one vote. The test instance is therefore classiﬁed as either y1 or y4 , depending on the tiebreaking procedure. ErrorCorrecting Output Coding A potential problem with the previous two approaches is that they are sensitive to the binary classiﬁcation errors. For the 1r approach given in Example 5.10,
307
Chapter 5
Classiﬁcation: Alternative Techniques
if at least of one of the binary classiﬁers makes a mistake in its prediction, then the ensemble may end up declaring a tie between classes or making a wrong prediction. For example, suppose the test instance is classiﬁed as (+, −, +, −) due to misclassiﬁcation by the third classiﬁer. In this case, it will be diﬃcult to tell whether the instance should be classiﬁed as y1 or y3 , unless the probability associated with each class prediction is taken into account. The errorcorrecting output coding (ECOC) method provides a more robust way for handling multiclass problems. The method is inspired by an informationtheoretic approach for sending messages across noisy channels. The idea behind this approach is to add redundancy into the transmitted message by means of a codeword, so that the receiver may detect errors in the received message and perhaps recover the original message if the number of errors is small. For multiclass learning, each class yi is represented by a unique bit string of length n known as its codeword. We then train n binary classiﬁers to predict each bit of the codeword string. The predicted class of a test instance is given by the codeword whose Hamming distance is closest to the codeword produced by the binary classiﬁers. Recall that the Hamming distance between a pair of bit strings is given by the number of bits that diﬀer. Example 5.11. Consider a multiclass problem where Y = {y1 , y2 , y3 , y4 }. Suppose we encode the classes using the following 7bit codewords: Class y1 y2 y3 y4
1 0 0 0
1 0 0 1
Codeword 1 1 1 0 0 1 1 1 0 0 1 0
1 1 0 1
1 1 1 0
Each bit of the codeword is used to train a binary classiﬁer. If a test instance is classiﬁed as (0,1,1,1,1,1,1) by the binary classiﬁers, then the Hamming distance between the codeword and y1 is 1, while the Hamming distance to the remaining classes is 3. The test instance is therefore classiﬁed as y1 . An interesting property of an errorcorrecting code is that if the minimum Hamming distance between any pair of codewords is d, then any (d − 1)/2) errors in the output code can be corrected using its nearest codeword. In Example 5.11, because the minimum Hamming distance between any pair of codewords is 4, the ensemble may tolerate errors made by one of the seven
308
5.9
Bibliographic Notes
binary classiﬁers. If there is more than one classiﬁer that makes a mistake, then the ensemble may not be able to compensate for the error. An important issue is how to design the appropriate set of codewords for diﬀerent classes. From coding theory, a vast number of algorithms have been developed for generating nbit codewords with bounded Hamming distance. However, the discussion of these algorithms is beyond the scope of this book. It is worthwhile mentioning that there is a signiﬁcant diﬀerence between the design of errorcorrecting codes for communication tasks compared to those used for multiclass learning. For communication, the codewords should maximize the Hamming distance between the rows so that error correction can be performed. Multiclass learning, however, requires that the rowwise and columnwise distances of the codewords must be well separated. A larger columnwise distance ensures that the binary classiﬁers are mutually independent, which is an important requirement for ensemble learning methods.
5.9
Bibliographic Notes
Mitchell [208] provides an excellent coverage on many classiﬁcation techniques from a machine learning perspective. Extensive coverage on classiﬁcation can also be found in Duda et al. [180], Webb [219], Fukunaga [187], Bishop [159], Hastie et al. [192], Cherkassky and Mulier [167], Witten and Frank [221], Hand et al. [190], Han and Kamber [189], and Dunham [181]. Direct methods for rulebased classiﬁers typically employ the sequential covering scheme for inducing classiﬁcation rules. Holte’s 1R [195] is the simplest form of a rulebased classiﬁer because its rule set contains only a single rule. Despite its simplicity, Holte found that for some data sets that exhibit a strong onetoone relationship between the attributes and the class label, 1R performs just as well as other classiﬁers. Other examples of rulebased classiﬁers include IREP [184], RIPPER [170], CN2 [168, 169], AQ [207], RISE [176], and ITRULE [214]. Table 5.9 shows a comparison of the characteristics of four of these classiﬁers. For rulebased classiﬁers, the rule antecedent can be generalized to include any propositional or ﬁrstorder logical expression (e.g., Horn clauses). Readers who are interested in ﬁrstorder logic rulebased classiﬁers may refer to references such as [208] or the vast literature on inductive logic programming [209]. Quinlan [211] proposed the C4.5rules algorithm for extracting classiﬁcation rules from decision trees. An indirect method for extracting rules from artiﬁcial neural networks was given by Andrews et al. in [157].
309
Chapter 5
Classiﬁcation: Alternative Techniques
Table 5.9. Comparison of various rulebased classifiers. RIPPER Rulegrowing strategy
Generaltospeciﬁc
CN2 (unordered) Generaltospeciﬁc
Evaluation Metric Stopping condition for rulegrowing Rule Pruning
FOIL’s Info gain
Laplace
All examples belong to the same class Reduced error pruning Positive and negative Error > 50% or based on MDL
No performance gain
Entropy and likelihood ratio No performance gain
None
None
None
Positive only
Positive only
No performance gain
No performance gain
Replace or modify rules Greedy
Statistical tests Beam search
None
Positive and negative All positive examples are covered None
Beam search
Beam search
Instance Elimination Stopping condition for adding rules Rule Set Pruning Search strategy
CN2 (ordered) Generaltospeciﬁc
AQR Generaltospeciﬁc (seeded by a positive example) Number of true positives Rules cover only positive class
Cover and Hart [172] presented an overview of the nearestneighbor classiﬁcation method from a Bayesian perspective. Aha provided both theoretical and empirical evaluations for instancebased methods in [155]. PEBLS, which was developed by Cost and Salzberg [171], is a nearestneighbor classiﬁcation algorithm that can handle data sets containing nominal attributes. Each training example in PEBLS is also assigned a weight factor that depends on the number of times the example helps make a correct prediction. Han et al. [188] developed a weightadjusted nearestneighbor algorithm, in which the feature weights are learned using a greedy, hillclimbing optimization algorithm. Na¨ıve Bayes classiﬁers have been investigated by many authors, including Langley et al. [203], Ramoni and Sebastiani [212], Lewis [204], and Domingos and Pazzani [178]. Although the independence assumption used in na¨ıve Bayes classiﬁers may seem rather unrealistic, the method has worked surprisingly well for applications such as text classiﬁcation. Bayesian belief networks provide a more ﬂexible approach by allowing some of the attributes to be interdependent. An excellent tutorial on Bayesian belief networks is given by Heckerman in [194]. Vapnik [217, 218] had written two authoritative books on Support Vector Machines (SVM). Other useful resources on SVM and kernel methods include the books by Cristianini and ShaweTaylor [173] and Sch¨ olkopf and Smola
310
5.9
Bibliographic Notes
[213]. There are several survey articles on SVM, including those written by Burges [164], Bennet et al. [158], Hearst [193], and Mangasarian [205]. A survey of ensemble methods in machine learning was given by Dietterich [174]. The bagging method was proposed by Breiman [161]. Freund and Schapire [186] developed the AdaBoost algorithm. Arcing, which stands for adaptive resampling and combining, is a variant of the boosting algorithm proposed by Breiman [162]. It uses the nonuniform weights assigned to training examples to resample the data for building an ensemble of training sets. Unlike AdaBoost, the votes of the base classiﬁers are not weighted when determining the class label of test examples. The random forest method was introduced by Breiman in [163]. Related work on mining rare and imbalanced data sets can be found in the survey papers written by Chawla et al. [166] and Weiss [220]. Samplingbased methods for mining imbalanced data sets have been investigated by many authors, such as Kubat and Matwin [202], Japkowitz [196], and Drummond and Holte [179]. Joshi et al. [199] discussed the limitations of boosting algorithms for rare class modeling. Other algorithms developed for mining rare classes include SMOTE [165], PNrule [198], and CREDOS [200]. Various alternative metrics that are wellsuited for class imbalanced problems are available. The precision, recall, and F1 measure are widely used metrics in information retrieval [216]. ROC analysis was originally used in signal detection theory. Bradley [160] investigated the use of area under the ROC curve as a performance metric for machine learning algorithms. A method for comparing classiﬁer performance using the convex hull of ROC curves was suggested by Provost and Fawcett in [210]. Ferri et al. [185] developed a methodology for performing ROC analysis on decision tree classiﬁers. They had also proposed a methodology for incorporating area under the ROC curve (AUC) as the splitting criterion during the treegrowing process. Joshi [197] examined the performance of these measures from the perspective of analyzing rare classes. A vast amount of literature on costsensitive learning can be found in the online proceedings of the ICML’2000 Workshop on costsensitive learning. The properties of a cost matrix had been studied by Elkan in [182]. Margineantu and Dietterich [206] examined various methods for incorporating cost information into the C4.5 learning algorithm, including wrapper methods, class distributionbased methods, and lossbased methods. Other costsensitive learning methods that are algorithmindependent include AdaCost [183], MetaCost [177], and costing [222].
311
Chapter 5
Classiﬁcation: Alternative Techniques
Extensive literature is also available on the subject of multiclass learning. This includes the works of Hastie and Tibshirani [191], Allwein et al. [156], Kong and Dietterich [201], and Tax and Duin [215]. The errorcorrecting output coding (ECOC) method was proposed by Dietterich and Bakiri [175]. They had also investigated techniques for designing codes that are suitable for solving multiclass problems.
Bibliography [155] D. W. Aha. A study of instancebased algorithms for supervised learning tasks: mathematical, empirical, and psychological evaluations. PhD thesis, University of California, Irvine, 1990. [156] E. L. Allwein, R. E. Schapire, and Y. Singer. Reducing Multiclass to Binary: A Unifying Approach to Margin Classiﬁers. Journal of Machine Learning Research, 1: 113–141, 2000. [157] R. Andrews, J. Diederich, and A. Tickle. A Survey and Critique of Techniques For Extracting Rules From Trained Artiﬁcial Neural Networks. Knowledge Based Systems, 8(6):373–389, 1995. [158] K. Bennett and C. Campbell. Support Vector Machines: Hype or Hallelujah. SIGKDD Explorations, 2(2):1–13, 2000. [159] C. M. Bishop. Neural Networks for Pattern Recognition. Oxford University Press, Oxford, U.K., 1995. [160] A. P. Bradley. The use of the area under the ROC curve in the Evaluation of Machine Learning Algorithms. Pattern Recognition, 30(7):1145–1149, 1997. [161] L. Breiman. Bagging Predictors. Machine Learning, 24(2):123–140, 1996. [162] L. Breiman. Bias, Variance, and Arcing Classiﬁers. Technical Report 486, University of California, Berkeley, CA, 1996. [163] L. Breiman. Random Forests. Machine Learning, 45(1):5–32, 2001. [164] C. J. C. Burges. A Tutorial on Support Vector Machines for Pattern Recognition. Data Mining and Knowledge Discovery, 2(2):121–167, 1998. [165] N. V. Chawla, K. W. Bowyer, L. O. Hall, and W. P. Kegelmeyer. SMOTE: Synthetic Minority Oversampling Technique. Journal of Artificial Intelligence Research, 16:321– 357, 2002. [166] N. V. Chawla, N. Japkowicz, and A. Kolcz. Editorial: Special Issue on Learning from Imbalanced Data Sets. SIGKDD Explorations, 6(1):1–6, 2004. [167] V. Cherkassky and F. Mulier. Learning from Data: Concepts, Theory, and Methods. Wiley Interscience, 1998. [168] P. Clark and R. Boswell. Rule Induction with CN2: Some Recent Improvements. In Machine Learning: Proc. of the 5th European Conf. (EWSL91), pages 151–163, 1991. [169] P. Clark and T. Niblett. The CN2 Induction Algorithm. Machine Learning, 3(4): 261–283, 1989. [170] W. W. Cohen. Fast Eﬀective Rule Induction. In Proc. of the 12th Intl. Conf. on Machine Learning, pages 115–123, Tahoe City, CA, July 1995. [171] S. Cost and S. Salzberg. A Weighted Nearest Neighbor Algorithm for Learning with Symbolic Features. Machine Learning, 10:57–78, 1993. [172] T. M. Cover and P. E. Hart. Nearest Neighbor Pattern Classiﬁcation. Knowledge Based Systems, 8(6):373–389, 1995.
312
Bibliography [173] N. Cristianini and J. ShaweTaylor. An Introduction to Support Vector Machines and Other Kernelbased Learning Methods. Cambridge University Press, 2000. [174] T. G. Dietterich. Ensemble Methods in Machine Learning. In First Intl. Workshop on Multiple Classifier Systems, Cagliari, Italy, 2000. [175] T. G. Dietterich and G. Bakiri. Solving Multiclass Learning Problems via ErrorCorrecting Output Codes. Journal of Artificial Intelligence Research, 2:263–286, 1995. [176] P. Domingos. The RISE system: Conquering without separating. In Proc. of the 6th IEEE Intl. Conf. on Tools with Artificial Intelligence, pages 704–707, New Orleans, LA, 1994. [177] P. Domingos. MetaCost: A General Method for Making Classiﬁers CostSensitive. In Proc. of the 5th Intl. Conf. on Knowledge Discovery and Data Mining, pages 155–164, San Diego, CA, August 1999. [178] P. Domingos and M. Pazzani. On the Optimality of the Simple Bayesian Classiﬁer under ZeroOne Loss. Machine Learning, 29(23):103–130, 1997. [179] C. Drummond and R. C. Holte. C4.5, Class imbalance, and Cost sensitivity: Why undersampling beats oversampling. In ICML’2004 Workshop on Learning from Imbalanced Data Sets II, Washington, DC, August 2003. [180] R. O. Duda, P. E. Hart, and D. G. Stork. Pattern Classification. John Wiley & Sons, Inc., New York, 2nd edition, 2001. [181] M. H. Dunham. Data Mining: Introductory and Advanced Topics. Prentice Hall, 2002. [182] C. Elkan. The Foundations of CostSensitive Learning. In Proc. of the 17th Intl. Joint Conf. on Artificial Intelligence, pages 973–978, Seattle, WA, August 2001. [183] W. Fan, S. J. Stolfo, J. Zhang, and P. K. Chan. AdaCost: misclassiﬁcation costsensitive boosting. In Proc. of the 16th Intl. Conf. on Machine Learning, pages 97–105, Bled, Slovenia, June 1999. [184] J. F¨ urnkranz and G. Widmer. Incremental reduced error pruning. In Proc. of the 11th Intl. Conf. on Machine Learning, pages 70–77, New Brunswick, NJ, July 1994. [185] C. Ferri, P. Flach, and J. HernandezOrallo. Learning Decision Trees Using the Area Under the ROC Curve. In Proc. of the 19th Intl. Conf. on Machine Learning, pages 139–146, Sydney, Australia, July 2002. [186] Y. Freund and R. E. Schapire. A decisiontheoretic generalization of online learning and an application to boosting. Journal of Computer and System Sciences, 55(1):119– 139, 1997. [187] K. Fukunaga. Introduction to Statistical Pattern Recognition. Academic Press, New York, 1990. [188] E.H. Han, G. Karypis, and V. Kumar. Text Categorization Using Weight Adjusted kNearest Neighbor Classiﬁcation. In Proc. of the 5th PacificAsia Conf. on Knowledge Discovery and Data Mining, Lyon, France, 2001. [189] J. Han and M. Kamber. Data Mining: Concepts and Techniques. Morgan Kaufmann Publishers, San Francisco, 2001. [190] D. J. Hand, H. Mannila, and P. Smyth. Principles of Data Mining. MIT Press, 2001. [191] T. Hastie and R. Tibshirani. Classiﬁcation by pairwise coupling. Annals of Statistics, 26(2):451–471, 1998. [192] T. Hastie, R. Tibshirani, and J. H. Friedman. The Elements of Statistical Learning: Data Mining, Inference, Prediction. Springer, New York, 2001. [193] M. Hearst. Trends & Controversies: Support Vector Machines. IEEE Intelligent Systems, 13(4):18–28, 1998.
313
Chapter 5
Classiﬁcation: Alternative Techniques
[194] D. Heckerman. Bayesian Networks for Data Mining. Data Mining and Knowledge Discovery, 1(1):79–119, 1997. [195] R. C. Holte. Very Simple Classiﬁcation Rules Perform Well on Most Commonly Used Data sets. Machine Learning, 11:63–91, 1993. [196] N. Japkowicz. The Class Imbalance Problem: Signiﬁcance and Strategies. In Proc. of the 2000 Intl. Conf. on Artificial Intelligence: Special Track on Inductive Learning, volume 1, pages 111–117, Las Vegas, NV, June 2000. [197] M. V. Joshi. On Evaluating Performance of Classiﬁers for Rare Classes. In Proc. of the 2002 IEEE Intl. Conf. on Data Mining, Maebashi City, Japan, December 2002. [198] M. V. Joshi, R. C. Agarwal, and V. Kumar. Mining Needles in a Haystack: Classifying Rare Classes via TwoPhase Rule Induction. In Proc. of 2001 ACMSIGMOD Intl. Conf. on Management of Data, pages 91–102, Santa Barbara, CA, June 2001. [199] M. V. Joshi, R. C. Agarwal, and V. Kumar. Predicting rare classes: can boosting make any weak learner strong? In Proc. of the 8th Intl. Conf. on Knowledge Discovery and Data Mining, pages 297–306, Edmonton, Canada, July 2002. [200] M. V. Joshi and V. Kumar. CREDOS: Classiﬁcation Using Ripple Down Structure (A Case for Rare Classes). In Proc. of the SIAM Intl. Conf. on Data Mining, pages 321–332, Orlando, FL, April 2004. [201] E. B. Kong and T. G. Dietterich. ErrorCorrecting Output Coding Corrects Bias and Variance. In Proc. of the 12th Intl. Conf. on Machine Learning, pages 313–321, Tahoe City, CA, July 1995. [202] M. Kubat and S. Matwin. Addressing the Curse of Imbalanced Training Sets: One Sided Selection. In Proc. of the 14th Intl. Conf. on Machine Learning, pages 179–186, Nashville, TN, July 1997. [203] P. Langley, W. Iba, and K. Thompson. An analysis of Bayesian classiﬁers. In Proc. of the 10th National Conf. on Artificial Intelligence, pages 223–228, 1992. [204] D. D. Lewis. Naive Bayes at Forty: The Independence Assumption in Information Retrieval. In Proc. of the 10th European Conf. on Machine Learning (ECML 1998), pages 4–15, 1998. [205] O. Mangasarian. Data Mining via Support Vector Machines. Technical Report Technical Report 0105, Data Mining Institute, May 2001. [206] D. D. Margineantu and T. G. Dietterich. Learning Decision Trees for Loss Minimization in MultiClass Problems. Technical Report 993003, Oregon State University, 1999. [207] R. S. Michalski, I. Mozetic, J. Hong, and N. Lavrac. The MultiPurpose Incremental Learning System AQ15 and Its Testing Application to Three Medical Domains. In Proc. of 5th National Conf. on Artificial Intelligence, Orlando, August 1986. [208] T. Mitchell. Machine Learning. McGrawHill, Boston, MA, 1997. [209] S. Muggleton. Foundations of Inductive Logic Programming. Prentice Hall, Englewood Cliﬀs, NJ, 1995. [210] F. J. Provost and T. Fawcett. Analysis and Visualization of Classiﬁer Performance: Comparison under Imprecise Class and Cost Distributions. In Proc. of the 3rd Intl. Conf. on Knowledge Discovery and Data Mining, pages 43–48, Newport Beach, CA, August 1997. [211] J. R. Quinlan. C4.5: Programs for Machine Learning. MorganKaufmann Publishers, San Mateo, CA, 1993. [212] M. Ramoni and P. Sebastiani. Robust Bayes classiﬁers. Artificial Intelligence, 125: 209–226, 2001.
314
5.10
Exercises
[213] B. Sch¨ olkopf and A. J. Smola. Learning with Kernels: Support Vector Machines, Regularization, Optimization, and Beyond. MIT Press, 2001. [214] P. Smyth and R. M. Goodman. An Information Theoretic Approach to Rule Induction from Databases. IEEE Trans. on Knowledge and Data Engineering, 4(4):301–316, 1992. [215] D. M. J. Tax and R. P. W. Duin. Using TwoClass Classiﬁers for Multiclass Classiﬁcation. In Proc. of the 16th Intl. Conf. on Pattern Recognition (ICPR 2002), pages 124–127, Quebec, Canada, August 2002. [216] C. J. van Rijsbergen. Information Retrieval. ButterworthHeinemann, Newton, MA, 1978. [217] V. Vapnik. The Nature of Statistical Learning Theory. Springer Verlag, New York, 1995. [218] V. Vapnik. Statistical Learning Theory. John Wiley & Sons, New York, 1998. [219] A. R. Webb. Statistical Pattern Recognition. John Wiley & Sons, 2nd edition, 2002. [220] G. M. Weiss. Mining with Rarity: A Unifying Framework. SIGKDD Explorations, 6 (1):7–19, 2004. [221] I. H. Witten and E. Frank. Data Mining: Practical Machine Learning Tools and Techniques with Java Implementations. Morgan Kaufmann, 1999. [222] B. Zadrozny, J. C. Langford, and N. Abe. CostSensitive Learning by CostProportionate Example Weighting. In Proc. of the 2003 IEEE Intl. Conf. on Data Mining, pages 435–442, Melbourne, FL, August 2003.
5.10
Exercises
1. Consider a binary classiﬁcation problem with the following set of attributes and attribute values: • Air Conditioner = {Working, Broken} • Engine = {Good, Bad} • Mileage = {High, Medium, Low} • Rust = {Yes, No} Suppose a rulebased classiﬁer produces the following rule set: Mileage = High −→ Value = Low Mileage = Low −→ Value = High Air Conditioner = Working, Engine = Good −→ Value = High Air Conditioner = Working, Engine = Bad −→ Value = Low Air Conditioner = Broken −→ Value = Low
(a) Are the rules mutually exclustive?
315
Chapter 5
Classiﬁcation: Alternative Techniques
(b) Is the rule set exhaustive? (c) Is ordering needed for this set of rules? (d) Do you need a default class for the rule set? 2. The RIPPER algorithm (by Cohen [170]) is an extension of an earlier algorithm called IREP (by F¨ urnkranz and Widmer [184]). Both algorithms apply the reducederror pruning method to determine whether a rule needs to be pruned. The reduced error pruning method uses a validation set to estimate the generalization error of a classiﬁer. Consider the following pair of rules: R1 : R2 :
A −→ C A ∧ B −→ C
R2 is obtained by adding a new conjunct, B, to the lefthand side of R1 . For this question, you will be asked to determine whether R2 is preferred over R1 from the perspectives of rulegrowing and rulepruning. To determine whether a rule should be pruned, IREP computes the following measure: vIREP =
p + (N − n) , P +N
where P is the total number of positive examples in the validation set, N is the total number of negative examples in the validation set, p is the number of positive examples in the validation set covered by the rule, and n is the number of negative examples in the validation set covered by the rule. vIREP is actually similar to classiﬁcation accuracy for the validation set. IREP favors rules that have higher values of vIREP . On the other hand, RIPPER applies the following measure to determine whether a rule should be pruned: vRIP P ER =
p−n . p+n
(a) Suppose R1 is covered by 350 positive examples and 150 negative examples, while R2 is covered by 300 positive examples and 50 negative examples. Compute the FOIL’s information gain for the rule R2 with respect to R1 . (b) Consider a validation set that contains 500 positive examples and 500 negative examples. For R1 , suppose the number of positive examples covered by the rule is 200, and the number of negative examples covered by the rule is 50. For R2 , suppose the number of positive examples covered by the rule is 100 and the number of negative examples is 5. Compute vIREP for both rules. Which rule does IREP prefer? (c) Compute vRIP P ER for the previous problem. Which rule does RIPPER prefer?
316
5.10
Exercises
3. C4.5rules is an implementation of an indirect method for generating rules from a decision tree. RIPPER is an implementation of a direct method for generating rules directly from data. (a) Discuss the strengths and weaknesses of both methods. (b) Consider a data set that has a large diﬀerence in the class size (i.e., some classes are much bigger than others). Which method (between C4.5rules and RIPPER) is better in terms of ﬁnding high accuracy rules for the small classes? 4. Consider a training set that contains 100 positive examples and 400 negative examples. For each of the following candidate rules, R1 : A −→ + (covers 4 positive and 1 negative examples), R2 : B −→ + (covers 30 positive and 10 negative examples), R3 : C −→ + (covers 100 positive and 90 negative examples), determine which is the best and worst candidate rule according to: (a) Rule accuracy. (b) FOIL’s information gain. (c) The likelihood ratio statistic. (d) The Laplace measure. (e) The mestimate measure (with k = 2 and p+ = 0.2). 5. Figure 5.4 illustrates the coverage of the classiﬁcation rules R1, R2, and R3. Determine which is the best and worst rule according to: (a) The likelihood ratio statistic. (b) The Laplace measure. (c) The mestimate measure (with k = 2 and p+ = 0.58). (d) The rule accuracy after R1 has been discovered, where none of the examples covered by R1 are discarded). (e) The rule accuracy after R1 has been discovered, where only the positive examples covered by R1 are discarded). (f) The rule accuracy after R1 has been discovered, where both positive and negative examples covered by R1 are discarded. 6.
(a) Suppose the fraction of undergraduate students who smoke is 15% and the fraction of graduate students who smoke is 23%. If oneﬁfth of the college students are graduate students and the rest are undergraduates, what is the probability that a student who smokes is a graduate student?
317
Chapter 5
Classiﬁcation: Alternative Techniques
(b) Given the information in part (a), is a randomly chosen college student more likely to be a graduate or undergraduate student? (c) Repeat part (b) assuming that the student is a smoker. (d) Suppose 30% of the graduate students live in a dorm but only 10% of the undergraduate students live in a dorm. If a student smokes and lives in the dorm, is he or she more likely to be a graduate or undergraduate student? You can assume independence between students who live in a dorm and those who smoke. 7. Consider the data set shown in Table 5.10 Table 5.10. Data set for Exercise 7. Record 1 2 3 4 5 6 7 8 9 10
A 0 0 0 0 0 1 1 1 1 1
B 0 0 1 1 0 0 0 0 1 0
C 0 1 1 1 1 1 1 1 1 1
Class + − − − + + − − + +
(a) Estimate the conditional probabilities for P (A+), P (B+), P (C+), P (A−), P (B−), and P (C−). (b) Use the estimate of conditional probabilities given in the previous question to predict the class label for a test sample (A = 0, B = 1, C = 0) using the na¨ıve Bayes approach. (c) Estimate the conditional probabilities using the mestimate approach, with p = 1/2 and m = 4. (d) Repeat part (b) using the conditional probabilities given in part (c). (e) Compare the two methods for estimating probabilities. Which method is better and why? 8. Consider the data set shown in Table 5.11. (a) Estimate the conditional probabilities for P (A = 1+), P (B = 1+), P (C = 1+), P (A = 1−), P (B = 1−), and P (C = 1−) using the same approach as in the previous problem.
318
5.10
Exercises
Table 5.11. Data set for Exercise 8. Instance 1 2 3 4 5 6 7 8 9 10
A 0 1 0 1 1 0 1 0 0 1
B 0 0 1 0 0 0 1 0 1 1
C 1 1 0 0 1 1 0 0 0 1
Class − + − − + + − − + +
(b) Use the conditional probabilities in part (a) to predict the class label for a test sample (A = 1, B = 1, C = 1) using the na¨ıve Bayes approach. (c) Compare P (A = 1), P (B = 1), and P (A = 1, B = 1). State the relationships between A and B. (d) Repeat the analysis in part (c) using P (A = 1), P (B = 0), and P (A = 1, B = 0). (e) Compare P (A = 1, B = 1Class = +) against P (A = 1Class = +) and P (B = 1Class = +). Are the variables conditionally independent given the class? 9.
(a) Explain how na¨ıve Bayes performs on the data set shown in Figure 5.46. (b) If each class is further divided such that there are four classes (A1, A2, B1, and B2), will na¨ıve Bayes perform better? (c) How will a decision tree perform on this data set (for the twoclass problem)? What if there are four classes?
10. Repeat the analysis shown in Example 5.3 for ﬁnding the location of a decision boundary using the following information: (a) The prior probabilities are P (Crocodile) = 2 × P (Alligator). (b) The prior probabilities are P (Alligator) = 2 × P (Crocodile). (c) The prior probabilities are the same, but their standard deviations are diﬀerent; i.e., σ(Crocodile) = 4 and σ(Alligator) = 2. 11. Figure 5.47 illustrates the Bayesian belief network for the data set shown in Table 5.12. (Assume that all the attributes are binary). (a) Draw the probability table for each node in the network.
319
Chapter 5
Classiﬁcation: Alternative Techniques Attributes
Distinguishing Attributes
Noise Attributes
A1 Class A A2 Records B1 Class B B2
Figure 5.46. Data set for Exercise 9.
Mileage
Air Conditioner
Engine Car Value
Figure 5.47. Bayesian belief network.
(b) Use the Bayesian network to compute P(Engine = Bad, Air Conditioner = Broken). 12. Given the Bayesian network shown in Figure 5.48, compute the following probabilities: (a) P (B = good, F = empty, G = empty, S = yes). (b) P (B = bad, F = empty, G = not empty, S = no). (c) Given that the battery is bad, compute the probability that the car will start. 13. Consider the onedimensional data set shown in Table 5.13.
320
5.10
Exercises
Table 5.12. Data set for Exercise 11. Mileage
Engine
Air Conditioner
Hi Hi Hi Hi Lo Lo Lo Lo
Good Good Bad Bad Good Good Bad Bad
Working Broken Working Broken Working Broken Working Broken
Number of Records with Car Value=Hi 3 1 1 0 9 5 1 0
P(B = bad) = 0.1
Number of Records with Car Value=Lo 4 2 5 4 0 1 2 2
P(F = empty) = 0.2
Battery
Fuel Gauge
P(G = empty  B = good, F = not empty) = 0.1 P(G = empty  B = good, F = empty) = 0.8 P(G = empty  B = bad, F = not empty) = 0.2 P(G = empty  B = bad, F = empty) = 0.9
Start P(S = no  B = good, F = not empty) = 0.1 P(S = no  B = good, F = empty) = 0.8 P(S = no  B = bad, F = not empty) = 0.9 P(S = no  B = bad, F = empty) = 1.0
Figure 5.48. Bayesian belief network for Exercise 12.
(a) Classify the data point x = 5.0 according to its 1, 3, 5, and 9nearest neighbors (using majority vote). (b) Repeat the previous analysis using the distanceweighted voting approach described in Section 5.2.1. 14. The nearestneighbor algorithm described in Section 5.2 can be extended to handle nominal attributes. A variant of the algorithm called PEBLS (Parallel ExamplarBased Learning System) by Cost and Salzberg [171] measures the distance between two values of a nominal attribute using the modiﬁed value diﬀerence metric (MVDM). Given a pair of nominal attribute values, V1 and
321
Chapter 5
x y
0.5 −
Classiﬁcation: Alternative Techniques Table 5.13. Data set for Exercise 13. 3.0 4.5 4.6 4.9 5.2 5.3 5.5 − + + + − − +
7.0 −
9.5 −
V2 , the distance between them is deﬁned as follows: k ni1 ni2 d(V1 , V2 ) = n1 − n2 , i=1
(5.84)
where nij is the number of examples from class i with attribute value Vj and nj is the number of examples with attribute value Vj . Consider the training set for the loan classiﬁcation problem shown in Figure 5.9. Use the MVDM measure to compute the distance between every pair of attribute values for the Home Owner and Marital Status attributes. 15. For each of the Boolean functions given below, state whether the problem is linearly separable. (a) A AND B AND C (b) NOT A AND B (c) (A OR B) AND (A OR C) (d) (A XOR B) AND (A OR B) 16.
(a) Demonstrate how the perceptron model can be used to represent the AND and OR functions between a pair of Boolean variables. (b) Comment on the disadvantage of using linear functions as activation functions for multilayer neural networks.
17. You are asked to evaluate the performance of two classiﬁcation models, M1 and M2 . The test set you have chosen contains 26 binary attributes, labeled as A through Z. Table 5.14 shows the posterior probabilities obtained by applying the models to the test set. (Only the posterior probabilities for the positive class are shown). As this is a twoclass problem, P (−) = 1 − P (+) and P (−A, . . . , Z) = 1−P (+A, . . . , Z). Assume that we are mostly interested in detecting instances from the positive class. (a) Plot the ROC curve for both M1 and M2 . (You should plot them on the same graph.) Which model do you think is better? Explain your reasons. (b) For model M1 , suppose you choose the cutoﬀ threshold to be t = 0.5. In other words, any test instances whose posterior probability is greater than t will be classiﬁed as a positive example. Compute the precision, recall, and Fmeasure for the model at this threshold value.
322
5.10
Exercises
Table 5.14. Posterior probabilities for Exercise 17. Instance 1 2 3 4 5 6 7 8 9 10
True Class + + − − + + − − + −
P (+A, . . . , Z, M1 ) 0.73 0.69 0.44 0.55 0.67 0.47 0.08 0.15 0.45 0.35
P (+A, . . . , Z, M2 ) 0.61 0.03 0.68 0.31 0.45 0.09 0.38 0.05 0.01 0.04
(c) Repeat the analysis for part (c) using the same cutoﬀ threshold on model M2 . Compare the F measure results for both models. Which model is better? Are the results consistent with what you expect from the ROC curve? (d) Repeat part (c) for model M1 using the threshold t = 0.1. Which threshold do you prefer, t = 0.5 or t = 0.1? Are the results consistent with what you expect from the ROC curve? 18. Following is a data set that contains two attributes, X and Y , and two class labels, “+” and “−”. Each attribute can take three diﬀerent values: 0, 1, or 2.
X 0 1 2 0 1 2 0 1 2
Y 0 0 0 1 1 1 2 2 2
Number of Instances + − 0 100 0 0 0 100 10 100 10 0 10 100 0 100 0 0 0 100
The concept for the “+” class is Y = 1 and the concept for the “−” class is X = 0 ∨ X = 2. (a) Build a decision tree on the data set. Does the tree capture the “+” and “−” concepts?
323
Chapter 5
Classiﬁcation: Alternative Techniques
(b) What are the accuracy, precision, recall, and F1 measure of the decision tree? (Note that precision, recall, and F1 measure are deﬁned with respect to the “+” class.) (c) Build a new decision tree with the following 0, 1, C(i, j) = Number of − instances Number of + instances ,
cost function: if i = j; if i = +, j = −; if i = −, j = +.
(Hint: only the leaves of the old decision tree need to be changed.) Does the decision tree capture the “+” concept? (d) What are the accuracy, precision, recall, and F1 measure of the new decision tree? 19.
(a) Consider the cost matrix for a twoclass problem. Let C(+, +) = C(−, −) = p, C(+, −) = C(−, +) = q, and q > p. Show that minimizing the cost function is equivalent to maximizing the classiﬁer’s accuracy. (b) Show that a cost matrix is scaleinvariant. For example, if the cost matrix is rescaled from C(i, j) −→ βC(i, j), where β is the scaling factor, the decision threshold (Equation 5.82) will remain unchanged. (c) Show that a cost matrix is translationinvariant. In other words, adding a constant factor to all entries in the cost matrix will not aﬀect the decision threshold (Equation 5.82).
20. Consider the task of building a classiﬁer from random data, where the attribute values are generated randomly irrespective of the class labels. Assume the data set contains records from two classes, “+” and “−.” Half of the data set is used for training while the remaining half is used for testing. (a) Suppose there are an equal number of positive and negative records in the data and the decision tree classiﬁer predicts every test record to be positive. What is the expected error rate of the classiﬁer on the test data? (b) Repeat the previous analysis assuming that the classiﬁer predicts each test record to be positive class with probability 0.8 and negative class with probability 0.2. (c) Suppose twothirds of the data belong to the positive class and the remaining onethird belong to the negative class. What is the expected error of a classiﬁer that predicts every test record to be positive? (d) Repeat the previous analysis assuming that the classiﬁer predicts each test record to be positive class with probability 2/3 and negative class with probability 1/3.
324
5.10
Exercises
21. Derive the dual Lagrangian for the linear SVM with nonseparable data where the objective function is 2 w2 +C ξi . 2 i=1 N
f (w) =
22. Consider the XOR problem where there are four training points: (1, 1, −), (1, 0, +), (0, 1, +), (0, 0, −). Transform the data into the following feature space: √ √ √ Φ = (1, 2x1 , 2x2 , 2x1 x2 , x21 , x22 ). Find the maximum margin linear decision boundary in the transformed space. 23. Given the data sets shown in Figures 5.49, explain how the decision tree, na¨ıve Bayes, and knearest neighbor classiﬁers would perform on these data sets.
325
Chapter 5
Classiﬁcation: Alternative Techniques
Attributes
Attributes Distinguishing Attributes
Noise Attributes
Distinguishing Attributes
Noise Attributes
Class A
Class A
Records
Records
Class B
Class B
(a) Synthetic data set 1.
(b) Synthetic data set 2.
Attributes
60% filled with 1
Noise Attributes
40% filled with 1
Class A
Records
40% filled with 1
60% filled with 1
Attribute Y
Distinguishing Distinguishing Attribute set 1 Attribute set 2
Class A
Class B
Class A
Class B
Class A
Class B
Class A
Class B
Class A
Class B
Class A
Class B
Class A
Class B
Class A
Class B
Class A
Class B
Class A
Class B
Class B
Attribute X
(c) Synthetic data set 3.
(d) Synthetic data set 4
Class A
Class B
Attribute Y
Attribute Y
Class A
Class B
Class B
Attribute X
Attribute X
(e) Synthetic data set 5.
(f) Synthetic data set 6.
Figure 5.49. Data set for Exercise 23.
326
6 Association Analysis: Basic Concepts and Algorithms Many business enterprises accumulate large quantities of data from their daytoday operations. For example, huge amounts of customer purchase data are collected daily at the checkout counters of grocery stores. Table 6.1 illustrates an example of such data, commonly known as market basket transactions. Each row in this table corresponds to a transaction, which contains a unique identiﬁer labeled T ID and a set of items bought by a given customer. Retailers are interested in analyzing the data to learn about the purchasing behavior of their customers. Such valuable information can be used to support a variety of businessrelated applications such as marketing promotions, inventory management, and customer relationship management. This chapter presents a methodology known as association analysis, which is useful for discovering interesting relationships hidden in large data sets. The uncovered relationships can be represented in the form of associaTable 6.1. An example of market basket transactions. T ID 1 2 3 4 5
Items {Bread, Milk} {Bread, Diapers, Beer, Eggs} {Milk, Diapers, Beer, Cola} {Bread, Milk, Diapers, Beer} {Bread, Milk, Diapers, Cola}
From Chapter 6 of Introduction to Data Mining, First Edition. PangNing Tan, Michael Steinbach, Vipin Kumar. Copyright © 2006 by Pearson Education, Inc. All rights reserved.
327
Chapter 6
Association Analysis
tion rules or sets of frequent items. For example, the following rule can be extracted from the data set shown in Table 6.1: {Diapers} −→ {Beer}. The rule suggests that a strong relationship exists between the sale of diapers and beer because many customers who buy diapers also buy beer. Retailers can use this type of rules to help them identify new opportunities for crossselling their products to the customers. Besides market basket data, association analysis is also applicable to other application domains such as bioinformatics, medical diagnosis, Web mining, and scientiﬁc data analysis. In the analysis of Earth science data, for example, the association patterns may reveal interesting connections among the ocean, land, and atmospheric processes. Such information may help Earth scientists develop a better understanding of how the diﬀerent elements of the Earth system interact with each other. Even though the techniques presented here are generally applicable to a wider variety of data sets, for illustrative purposes, our discussion will focus mainly on market basket data. There are two key issues that need to be addressed when applying association analysis to market basket data. First, discovering patterns from a large transaction data set can be computationally expensive. Second, some of the discovered patterns are potentially spurious because they may happen simply by chance. The remainder of this chapter is organized around these two issues. The ﬁrst part of the chapter is devoted to explaining the basic concepts of association analysis and the algorithms used to eﬃciently mine such patterns. The second part of the chapter deals with the issue of evaluating the discovered patterns in order to prevent the generation of spurious results.
6.1
Problem Deﬁnition
This section reviews the basic terminology used in association analysis and presents a formal description of the task. Binary Representation Market basket data can be represented in a binary format as shown in Table 6.2, where each row corresponds to a transaction and each column corresponds to an item. An item can be treated as a binary variable whose value is one if the item is present in a transaction and zero otherwise. Because the presence of an item in a transaction is often considered more important than its absence, an item is an asymmetric binary variable.
328
6.1
Problem Deﬁnition
Table 6.2. A binary 0/1 representation of market basket data. TID 1 2 3 4 5
Bread 1 1 0 1 1
Milk 1 0 1 1 1
Diapers 0 1 1 1 1
Beer 0 1 1 1 0
Eggs 0 1 0 0 0
Cola 0 0 1 0 1
This representation is perhaps a very simplistic view of real market basket data because it ignores certain important aspects of the data such as the quantity of items sold or the price paid to purchase them. Methods for handling such nonbinary data will be explained in Chapter 7. Itemset and Support Count Let I = {i1 ,i2 ,. . .,id } be the set of all items in a market basket data and T = {t1 , t2 , . . . , tN } be the set of all transactions. Each transaction ti contains a subset of items chosen from I. In association analysis, a collection of zero or more items is termed an itemset. If an itemset contains k items, it is called a kitemset. For instance, {Beer, Diapers, Milk} is an example of a 3itemset. The null (or empty) set is an itemset that does not contain any items. The transaction width is deﬁned as the number of items present in a transaction. A transaction tj is said to contain an itemset X if X is a subset of tj . For example, the second transaction shown in Table 6.2 contains the itemset {Bread, Diapers} but not {Bread, Milk}. An important property of an itemset is its support count, which refers to the number of transactions that contain a particular itemset. Mathematically, the support count, σ(X), for an itemset X can be stated as follows: σ(X) = {ti X ⊆ ti , ti ∈ T }, where the symbol  ·  denote the number of elements in a set. In the data set shown in Table 6.2, the support count for {Beer, Diapers, Milk} is equal to two because there are only two transactions that contain all three items. Association Rule An association rule is an implication expression of the form X −→ Y , where X and Y are disjoint itemsets, i.e., X ∩ Y = ∅. The strength of an association rule can be measured in terms of its support and conﬁdence. Support determines how often a rule is applicable to a given
329
Chapter 6
Association Analysis
data set, while conﬁdence determines how frequently items in Y appear in transactions that contain X. The formal deﬁnitions of these metrics are Support, s(X −→ Y ) = Conﬁdence, c(X −→ Y ) =
σ(X ∪ Y ) ; N σ(X ∪ Y ) . σ(X)
(6.1) (6.2)
Example 6.1. Consider the rule {Milk, Diapers} −→ {Beer}. Since the support count for {Milk, Diapers, Beer} is 2 and the total number of transactions is 5, the rule’s support is 2/5 = 0.4. The rule’s conﬁdence is obtained by dividing the support count for {Milk, Diapers, Beer} by the support count for {Milk, Diapers}. Since there are 3 transactions that contain milk and diapers, the conﬁdence for this rule is 2/3 = 0.67. Why Use Support and Conﬁdence? Support is an important measure because a rule that has very low support may occur simply by chance. A low support rule is also likely to be uninteresting from a business perspective because it may not be proﬁtable to promote items that customers seldom buy together (with the exception of the situation described in Section 6.8). For these reasons, support is often used to eliminate uninteresting rules. As will be shown in Section 6.2.1, support also has a desirable property that can be exploited for the eﬃcient discovery of association rules. Conﬁdence, on the other hand, measures the reliability of the inference made by a rule. For a given rule X −→ Y , the higher the conﬁdence, the more likely it is for Y to be present in transactions that contain X. Conﬁdence also provides an estimate of the conditional probability of Y given X. Association analysis results should be interpreted with caution. The inference made by an association rule does not necessarily imply causality. Instead, it suggests a strong cooccurrence relationship between items in the antecedent and consequent of the rule. Causality, on the other hand, requires knowledge about the causal and eﬀect attributes in the data and typically involves relationships occurring over time (e.g., ozone depletion leads to global warming). Formulation of Association Rule Mining Problem rule mining problem can be formally stated as follows:
The association
Deﬁnition 6.1 (Association Rule Discovery). Given a set of transactions T , ﬁnd all the rules having support ≥ minsup and conﬁdence ≥ minconf , where minsup and minconf are the corresponding support and conﬁdence thresholds.
330
6.1
Problem Deﬁnition
A bruteforce approach for mining association rules is to compute the support and conﬁdence for every possible rule. This approach is prohibitively expensive because there are exponentially many rules that can be extracted from a data set. More speciﬁcally, the total number of possible rules extracted from a data set that contains d items is R = 3d − 2d+1 + 1.
(6.3)
The proof for this equation is left as an exercise to the readers (see Exercise 5 on page 405). Even for the small data set shown in Table 6.1, this approach requires us to compute the support and conﬁdence for 36 − 27 + 1 = 602 rules. More than 80% of the rules are discarded after applying minsup = 20% and minconf = 50%, thus making most of the computations become wasted. To avoid performing needless computations, it would be useful to prune the rules early without having to compute their support and conﬁdence values. An initial step toward improving the performance of association rule mining algorithms is to decouple the support and conﬁdence requirements. From Equation 6.2, notice that the support of a rule X −→ Y depends only on the support of its corresponding itemset, X ∪ Y . For example, the following rules have identical support because they involve items from the same itemset, {Beer, Diapers, Milk}: {Beer, Diapers} −→ {Milk}, {Diapers, Milk} −→ {Beer}, {Milk} −→ {Beer,Diapers},
{Beer, Milk} −→ {Diapers}, {Beer} −→ {Diapers, Milk}, {Diapers} −→ {Beer,Milk}.
If the itemset is infrequent, then all six candidate rules can be pruned immediately without our having to compute their conﬁdence values. Therefore, a common strategy adopted by many association rule mining algorithms is to decompose the problem into two major subtasks: 1. Frequent Itemset Generation, whose objective is to ﬁnd all the itemsets that satisfy the minsup threshold. These itemsets are called frequent itemsets. 2. Rule Generation, whose objective is to extract all the highconﬁdence rules from the frequent itemsets found in the previous step. These rules are called strong rules. The computational requirements for frequent itemset generation are generally more expensive than those of rule generation. Eﬃcient techniques for generating frequent itemsets and association rules are discussed in Sections 6.2 and 6.3, respectively.
331
Chapter 6
Association Analysis null
a
b
c
d
e
ab
ac
ad
ae
bc
bd
be
cd
ce
de
abc
abd
abe
acd
ace
ade
bcd
bce
bde
cde
abcd
abce
abde
acde
bcde
abcde
Figure 6.1. An itemset lattice.
6.2
Frequent Itemset Generation
A lattice structure can be used to enumerate the list of all possible itemsets. Figure 6.1 shows an itemset lattice for I = {a, b, c, d, e}. In general, a data set that contains k items can potentially generate up to 2k − 1 frequent itemsets, excluding the null set. Because k can be very large in many practical applications, the search space of itemsets that need to be explored is exponentially large. A bruteforce approach for ﬁnding frequent itemsets is to determine the support count for every candidate itemset in the lattice structure. To do this, we need to compare each candidate against every transaction, an operation that is shown in Figure 6.2. If the candidate is contained in a transaction, its support count will be incremented. For example, the support for {Bread, Milk} is incremented three times because the itemset is contained in transactions 1, 4, and 5. Such an approach can be very expensive because it requires O(N M w) comparisons, where N is the number of transactions, M = 2k − 1 is the number of candidate itemsets, and w is the maximum transaction width.
332
6.2
Frequent Itemset Generation Candidates
TID 1 2 N 3 4 5
Transactions Items Bread, Milk Bread, Diapers, Beer, Eggs Milk, Diapers, Beer, Coke Bread, Milk, Diapers, Beer Bread, Milk, Diapers, Coke
M
Figure 6.2. Counting the support of candidate itemsets.
There are several ways to reduce the computational complexity of frequent itemset generation. 1. Reduce the number of candidate itemsets (M ). The Apriori principle, described in the next section, is an eﬀective way to eliminate some of the candidate itemsets without counting their support values. 2. Reduce the number of comparisons. Instead of matching each candidate itemset against every transaction, we can reduce the number of comparisons by using more advanced data structures, either to store the candidate itemsets or to compress the data set. We will discuss these strategies in Sections 6.2.4 and 6.6.
6.2.1
The Apriori Principle
This section describes how the support measure helps to reduce the number of candidate itemsets explored during frequent itemset generation. The use of support for pruning candidate itemsets is guided by the following principle. Theorem 6.1 (Apriori Principle). If an itemset is frequent, then all of its subsets must also be frequent. To illustrate the idea behind the Apriori principle, consider the itemset lattice shown in Figure 6.3. Suppose {c, d, e} is a frequent itemset. Clearly, any transaction that contains {c, d, e} must also contain its subsets, {c, d}, {c, e}, {d, e}, {c}, {d}, and {e}. As a result, if {c, d, e} is frequent, then all subsets of {c, d, e} (i.e., the shaded itemsets in this ﬁgure) must also be frequent.
333
Chapter 6
Association Analysis null
b
a
c
d
e
ab
ac
ad
ae
bc
bd
be
cd
ce
de
abc
abd
abe
acd
ace
ade
bcd
bce
bde
cde
abcd
abce
abde
acde
bcde
Frequent Itemset abcde
Figure 6.3. An illustration of the Apriori principle. If {c, d, e} is frequent, then all subsets of this itemset are frequent.
Conversely, if an itemset such as {a, b} is infrequent, then all of its supersets must be infrequent too. As illustrated in Figure 6.4, the entire subgraph containing the supersets of {a, b} can be pruned immediately once {a, b} is found to be infrequent. This strategy of trimming the exponential search space based on the support measure is known as supportbased pruning. Such a pruning strategy is made possible by a key property of the support measure, namely, that the support for an itemset never exceeds the support for its subsets. This property is also known as the antimonotone property of the support measure. Deﬁnition 6.2 (Monotonicity Property). Let I be a set of items, and J = 2I be the power set of I. A measure f is monotone (or upward closed) if ∀X, Y ∈ J : (X ⊆ Y ) −→ f (X) ≤ f (Y ),
334
6.2
Frequent Itemset Generation
null
Infrequent Itemset
a
b
c
d
e
ab
ac
ad
ae
bc
bd
be
cd
ce
de
abc
abd
abe
acd
ace
ade
bcd
bce
bde
cde
abcd
abce
abde
acde
bcde
Pruned Supersets abcde
Figure 6.4. An illustration of supportbased pruning. If {a, b} is infrequent, then all supersets of {a, b} are infrequent.
which means that if X is a subset of Y , then f (X) must not exceed f (Y ). On the other hand, f is antimonotone (or downward closed) if ∀X, Y ∈ J : (X ⊆ Y ) −→ f (Y ) ≤ f (X), which means that if X is a subset of Y , then f (Y ) must not exceed f (X). Any measure that possesses an antimonotone property can be incorporated directly into the mining algorithm to eﬀectively prune the exponential search space of candidate itemsets, as will be shown in the next section.
6.2.2
Frequent Itemset Generation in the Apriori Algorithm
Apriori is the ﬁrst association rule mining algorithm that pioneered the use of supportbased pruning to systematically control the exponential growth of candidate itemsets. Figure 6.5 provides a highlevel illustration of the frequent itemset generation part of the Apriori algorithm for the transactions shown in
335
Chapter 6
Association Analysis
Candidate 1Itemsets Item Count Beer 3 4 Bread Cola 2 Diapers 4 4 Milk Eggs 1
Itemsets removed because of low support
Minimum support count = 3
Candidate 2Itemsets Itemset Count {Beer, Bread} 2 {Beer, Diapers} 3 {Beer, Milk} 2 {Bread, Diapers} 3 {Bread, Milk} 3 {Diapers, Milk} 3 Candidate 3Itemsets Itemset Count {Bread, Diapers, Milk} 3
Figure 6.5. Illustration of frequent itemset generation using the Apriori algorithm.
Table 6.1. We assume that the support threshold is 60%, which is equivalent to a minimum support count equal to 3. Initially, every item is considered as a candidate 1itemset. After counting their supports, the candidate itemsets {Cola} and {Eggs} are discarded because they appear in fewer than three transactions. In the next iteration, candidate 2itemsets are generated using only the frequent 1itemsets because the Apriori principle ensures that all supersets of the infrequent 1itemsets must be infrequent. Because there are only four frequent 1itemsets, the number of candidate 2itemsets generated by the algorithm is 42 = 6. Two of these six candidates, {Beer, Bread} and {Beer, Milk}, are subsequently found to be infrequent after computing their support values. The remaining four candidates are frequent, and thus will be used to candidate 6generate 3itemsets. Without supportbased pruning, there are 3 = 20 candidate 3itemsets that can be formed using the six items given in this example. With the Apriori principle, we only need to keep candidate 3itemsets whose subsets are frequent. The only candidate that has this property is {Bread, Diapers, Milk}. The eﬀectiveness of the Apriori pruning strategy can be shown by counting the number of candidate itemsets generated. A bruteforce strategy of
336
6.2
Frequent Itemset Generation
enumerating all itemsets (up to size 3) as candidates will produce 6 6 6 + + = 6 + 15 + 20 = 41 1 2 3 candidates. With the Apriori principle, this number decreases to 6 4 + + 1 = 6 + 6 + 1 = 13 1 2 candidates, which represents a 68% reduction in the number of candidate itemsets even in this simple example. The pseudocode for the frequent itemset generation part of the Apriori algorithm is shown in Algorithm 6.1. Let Ck denote the set of candidate kitemsets and Fk denote the set of frequent kitemsets: • The algorithm initially makes a single pass over the data set to determine the support of each item. Upon completion of this step, the set of all frequent 1itemsets, F1 , will be known (steps 1 and 2). • Next, the algorithm will iteratively generate new candidate kitemsets using the frequent (k − 1)itemsets found in the previous iteration (step 5). Candidate generation is implemented using a function called apriorigen, which is described in Section 6.2.3. Algorithm 6.1 Frequent itemset generation of the Apriori algorithm. 1: 2: 3: 4: 5: 6: 7: 8: 9: 10: 11: 12: 13: 14:
k = 1. Fk = { i  i ∈ I ∧ σ({i}) ≥ N × minsup}. {Find all frequent 1itemsets} repeat k = k + 1. Ck = apriorigen(Fk−1 ). {Generate candidate itemsets} for each transaction t ∈ T do Ct = subset(Ck , t). {Identify all candidates that belong to t} for each candidate itemset c ∈ Ct do σ(c) = σ(c) + 1. {Increment support count} end for end for Fk = { c  c ∈ Ck ∧ σ(c) ≥ N × minsup}. {Extract the frequent kitemsets} until Fk = ∅ Result = Fk .
337
Chapter 6
Association Analysis
• To count the support of the candidates, the algorithm needs to make an additional pass over the data set (steps 6–10). The subset function is used to determine all the candidate itemsets in Ck that are contained in each transaction t. The implementation of this function is described in Section 6.2.4. • After counting their supports, the algorithm eliminates all candidate itemsets whose support counts are less than minsup (step 12). • The algorithm terminates when there are no new frequent itemsets generated, i.e., Fk = ∅ (step 13). The frequent itemset generation part of the Apriori algorithm has two important characteristics. First, it is a levelwise algorithm; i.e., it traverses the itemset lattice one level at a time, from frequent 1itemsets to the maximum size of frequent itemsets. Second, it employs a generateandtest strategy for ﬁnding frequent itemsets. At each iteration, new candidate itemsets are generated from the frequent itemsets found in the previous iteration. The support for each candidate is then counted and tested against the minsup threshold. The total number of iterations needed by the algorithm is kmax + 1, where kmax is the maximum size of the frequent itemsets.
6.2.3
Candidate Generation and Pruning
The apriorigen function shown in Step 5 of Algorithm 6.1 generates candidate itemsets by performing the following two operations: 1. Candidate Generation. This operation generates new candidate kitemsets based on the frequent (k − 1)itemsets found in the previous iteration. 2. Candidate Pruning. This operation eliminates some of the candidate kitemsets using the supportbased pruning strategy. To illustrate the candidate pruning operation, consider a candidate kitemset, X = {i1 , i2 , . . . , ik }. The algorithm must determine whether all of its proper subsets, X − {ij } (∀j = 1, 2, . . . , k), are frequent. If one of them is infrequent, then X is immediately pruned. This approach can eﬀectively reduce the number of candidate itemsets considered during support counting. The complexity of this operation is O(k) for each candidate kitemset. However, as will be shown later, we do not have to examine all k subsets of a given candidate itemset. If m of the k subsets were used to generate a candidate, we only need to check the remaining k − m subsets during candidate pruning.
338
6.2
Frequent Itemset Generation
In principle, there are many ways to generate candidate itemsets. The following is a list of requirements for an eﬀective candidate generation procedure: 1. It should avoid generating too many unnecessary candidates. A candidate itemset is unnecessary if at least one of its subsets is infrequent. Such a candidate is guaranteed to be infrequent according to the antimonotone property of support. 2. It must ensure that the candidate set is complete, i.e., no frequent itemsets are left out by the candidate generation procedure. To ensure completeness, the set of candidate itemsets must subsume the set of all frequent itemsets, i.e., ∀k : Fk ⊆ Ck . 3. It should not generate the same candidate itemset more than once. For example, the candidate itemset {a, b, c, d} can be generated in many ways—by merging {a, b, c} with {d}, {b, d} with {a, c}, {c} with {a, b, d}, etc. Generation of duplicate candidates leads to wasted computations and thus should be avoided for eﬃciency reasons. Next, we will brieﬂy describe several candidate generation procedures, including the one used by the apriorigen function. BruteForce Method The bruteforce method considers every kitemset as a potential candidate and then applies the candidate pruning step to remove any unnecessary candidates (see Figure 6.6). The number of candidate itemsets generated at level k is equal to kd , where d is the total number of items. Although candidate generation is rather trivial, candidate pruning becomes extremely expensive because a large number of itemsets must be examined. Given that the amount of computations needed is O(k), d candidate d for each = O d · 2d−1 . the overall complexity of this method is O k=1 k × k Fk−1 × F1 Method An alternative method for candidate generation is to extend each frequent (k − 1)itemset with other frequent items. Figure 6.7 illustrates how a frequent 2itemset such as {Beer, Diapers} can be augmented with a frequent item such as Bread to produce a candidate 3itemset {Beer, Diapers, Bread}. This method will produce O(Fk−1  × F1 ) candidate kitemsets, where Fj  isthe number of frequent jitemsets. The overall complexity of this step is O( k kFk−1 F1 ). The procedure is complete because every frequent kitemset is composed of a frequent (k − 1)itemset and a frequent 1itemset. Therefore, all frequent kitemsets are part of the candidate kitemsets generated by this procedure.
339
Chapter 6
Association Analysis Candidate Generation
Items Item Beer Bread Cola Diapers Milk Eggs
Itemset {Beer, Bread, Cola} {Beer, Bread, Diapers} {Beer, Bread, Milk} {Beer, Bread, Eggs} {Beer, Cola, Diapers} {Beer, Cola, Milk} {Beer, Cola, Eggs} {Beer, Diapers, Milk} {Beer, Diapers, Eggs} {Beer, Milk, Eggs} {Bread, Cola, Diapers} {Bread, Cola, Milk} {Bread, Cola, Eggs} {Bread, Diapers, Milk} {Bread, Diapers, Eggs} {Bread, Milk, Eggs} {Cola, Diapers, Milk} {Cola, Diapers, Eggs} {Cola, Milk, Eggs} {Diapers, Milk, Eggs}
Candidate Pruning Itemset {Bread, Diapers, Milk}
Figure 6.6. A bruteforce method for generating candidate 3itemsets. Frequent 2itemset Itemset {Beer, Diapers} {Bread, Diapers} {Bread, Milk} {Diapers, Milk} Candidate Generation
Frequent 1itemset Item Beer Bread Diapers Milk
Itemset {Beer, Diapers, Bread} {Beer, Diapers, Milk} {Bread, Diapers, Milk} {Bread, Milk, Beer}
Candidate Pruning Itemset {Bread, Diapers, Milk}
Figure 6.7. Generating and pruning candidate kitemsets by merging a frequent (k − 1)itemset with a frequent item. Note that some of the candidates are unnecessary because their subsets are infrequent.
This approach, however, does not prevent the same candidate itemset from being generated more than once. For instance, {Bread, Diapers, Milk} can be generated by merging {Bread, Diapers} with {Milk}, {Bread, Milk} with {Diapers}, or {Diapers, Milk} with {Bread}. One way to avoid generating
340
6.2
Frequent Itemset Generation
duplicate candidates is by ensuring that the items in each frequent itemset are kept sorted in their lexicographic order. Each frequent (k−1)itemset X is then extended with frequent items that are lexicographically larger than the items in X. For example, the itemset {Bread, Diapers} can be augmented with {Milk} since Milk is lexicographically larger than Bread and Diapers. However, we should not augment {Diapers, Milk} with {Bread} nor {Bread, Milk} with {Diapers} because they violate the lexicographic ordering condition. While this procedure is a substantial improvement over the bruteforce method, it can still produce a large number of unnecessary candidates. For example, the candidate itemset obtained by merging {Beer, Diapers} with {Milk} is unnecessary because one of its subsets, {Beer, Milk}, is infrequent. There are several heuristics available to reduce the number of unnecessary candidates. For example, note that, for every candidate kitemset that survives the pruning step, every item in the candidate must be contained in at least k − 1 of the frequent (k − 1)itemsets. Otherwise, the candidate is guaranteed to be infrequent. For example, {Beer, Diapers, Milk} is a viable candidate 3itemset only if every item in the candidate, including Beer, is contained in at least two frequent 2itemsets. Since there is only one frequent 2itemset containing Beer, all candidate itemsets involving Beer must be infrequent. Fk−1 ×Fk−1 Method The candidate generation procedure in the apriorigen function merges a pair of frequent (k − 1)itemsets only if their ﬁrst k − 2 items are identical. Let A = {a1 , a2 , . . . , ak−1 } and B = {b1 , b2 , . . . , bk−1 } be a pair of frequent (k − 1)itemsets. A and B are merged if they satisfy the following conditions: ai = bi (for i = 1, 2, . . . , k − 2) and ak−1 = bk−1 . In Figure 6.8, the frequent itemsets {Bread, Diapers} and {Bread, Milk} are merged to form a candidate 3itemset {Bread, Diapers, Milk}. The algorithm does not have to merge {Beer, Diapers} with {Diapers, Milk} because the ﬁrst item in both itemsets is diﬀerent. Indeed, if {Beer, Diapers, Milk} is a viable candidate, it would have been obtained by merging {Beer, Diapers} with {Beer, Milk} instead. This example illustrates both the completeness of the candidate generation procedure and the advantages of using lexicographic ordering to prevent duplicate candidates. However, because each candidate is obtained by merging a pair of frequent (k−1)itemsets, an additional candidate pruning step is needed to ensure that the remaining k − 2 subsets of the candidate are frequent.
341
Chapter 6
Association Analysis
Frequent 2itemset Itemset {Beer, Diapers} {Bread, Diapers} {Bread, Milk} {Diapers, Milk}
Frequent 2itemset
Candidate Generation
Candidate Pruning
Itemset {Bread, Diapers, Milk}
Itemset {Bread, Diapers, Milk}
Itemset {Beer, Diapers} {Bread, Diapers} {Bread, Milk} {Diapers, Milk}
Figure 6.8. Generating and pruning candidate kitemsets by merging pairs of frequent (k−1)itemsets.
6.2.4
Support Counting
Support counting is the process of determining the frequency of occurrence for every candidate itemset that survives the candidate pruning step of the apriorigen function. Support counting is implemented in steps 6 through 11 of Algorithm 6.1. One approach for doing this is to compare each transaction against every candidate itemset (see Figure 6.2) and to update the support counts of candidates contained in the transaction. This approach is computationally expensive, especially when the numbers of transactions and candidate itemsets are large. An alternative approach is to enumerate the itemsets contained in each transaction and use them to update the support counts of their respective candidate itemsets. To illustrate, 5 consider a transaction t that contains ﬁve items, {1, 2, 3, 5, 6}. There are 3 = 10 itemsets of size 3 contained in this transaction. Some of the itemsets may correspond to the candidate 3itemsets under investigation, in which case, their support counts are incremented. Other subsets of t that do not correspond to any candidates can be ignored. Figure 6.9 shows a systematic way for enumerating the 3itemsets contained in t. Assuming that each itemset keeps its items in increasing lexicographic order, an itemset can be enumerated by specifying the smallest item ﬁrst, followed by the larger items. For instance, given t = {1, 2, 3, 5, 6}, all the 3itemsets contained in t must begin with item 1, 2, or 3. It is not possible to construct a 3itemset that begins with items 5 or 6 because there are only two
342
6.2
Frequent Itemset Generation
Transaction, t 1 2 3 5 6
Level 1 1 2 3 5 6
2 3 5 6
3 5 6
Level 2 12 3 5 6
13 5 6
123 125 126
135 136
Level 3
15 6
156
23 5 6
235 236
25 6
256
35 6
356
Subsets of 3 items
Figure 6.9. Enumerating subsets of three items from a transaction t.
items in t whose labels are greater than or equal to 5. The number of ways to specify the ﬁrst item of a 3itemset contained in t is illustrated by the Level 1 preﬁx structures depicted in Figure 6.9. For instance, 1 2 3 5 6 represents a 3itemset that begins with item 1, followed by two more items chosen from the set {2, 3, 5, 6}. After ﬁxing the ﬁrst item, the preﬁx structures at Level 2 represent the number of ways to select the second item. For example, 1 2 3 5 6 corresponds to itemsets that begin with preﬁx (1 2) and are followed by items 3, 5, or 6. Finally, the preﬁx structures at Level 3 represent the complete set of 3itemsets contained in t. For example, the 3itemsets that begin with preﬁx {1 2} are {1, 2, 3}, {1, 2, 5}, and {1, 2, 6}, while those that begin with preﬁx {2 3} are {2, 3, 5} and {2, 3, 6}. The preﬁx structures shown in Figure 6.9 demonstrate how itemsets contained in a transaction can be systematically enumerated, i.e., by specifying their items one by one, from the leftmost item to the rightmost item. We still have to determine whether each enumerated 3itemset corresponds to an existing candidate itemset. If it matches one of the candidates, then the support count of the corresponding candidate is incremented. In the next section, we illustrate how this matching operation can be performed eﬃciently using a hash tree structure.
343
Chapter 6
Association Analysis Hash Tree
Leaf nodes {Beer, Bread} containing candidate {Beer, Diapers} {Beer, Milk} 2itemsets
{Bread, Diapers} {Diapers, Milk} {Bread, Milk}
Transactions
TID 1 2 3 4 5
Items Bread, Milk Bread, Diapers, Beer, Eggs Milk, Diapers, Beer, Cola Bread, Milk, Diapers, Beer Bread, Milk, Diapers, Cola
Figure 6.10. Counting the support of itemsets using hash structure.
Support Counting Using a Hash Tree In the Apriori algorithm, candidate itemsets are partitioned into diﬀerent buckets and stored in a hash tree. During support counting, itemsets contained in each transaction are also hashed into their appropriate buckets. That way, instead of comparing each itemset in the transaction with every candidate itemset, it is matched only against candidate itemsets that belong to the same bucket, as shown in Figure 6.10. Figure 6.11 shows an example of a hash tree structure. Each internal node of the tree uses the following hash function, h(p) = p mod 3, to determine which branch of the current node should be followed next. For example, items 1, 4, and 7 are hashed to the same branch (i.e., the leftmost branch) because they have the same remainder after dividing the number by 3. All candidate itemsets are stored at the leaf nodes of the hash tree. The hash tree shown in Figure 6.11 contains 15 candidate 3itemsets, distributed across 9 leaf nodes. Consider a transaction, t = {1, 2, 3, 5, 6}. To update the support counts of the candidate itemsets, the hash tree must be traversed in such a way that all the leaf nodes containing candidate 3itemsets belonging to t must be visited at least once. Recall that the 3itemsets contained in t must begin with items 1, 2, or 3, as indicated by the Level 1 preﬁx structures shown in Figure 6.9. Therefore, at the root node of the hash tree, the items 1, 2, and 3 of the transaction are hashed separately. Item 1 is hashed to the left child of the root node, item 2 is hashed to the middle child, and item 3 is hashed to the right child. At the next level of the tree, the transaction is hashed on the second
344
6.2
Frequent Itemset Generation
Hash Function
1,4,7
3,6,9 2,5,8
1+ 2356
Transaction
2+
356
12356
3+ 56
Candidate Hash Tree
234 567 145
136
345
356
367
357
368
689 124
125
457
458
159
Figure 6.11. Hashing a transaction at the root node of a hash tree.
item listed in the Level 2 structures shown in Figure 6.9. For example, after hashing on item 1 at the root node, items 2, 3, and 5 of the transaction are hashed. Items 2 and 5 are hashed to the middle child, while item 3 is hashed to the right child, as shown in Figure 6.12. This process continues until the leaf nodes of the hash tree are reached. The candidate itemsets stored at the visited leaf nodes are compared against the transaction. If a candidate is a subset of the transaction, its support count is incremented. In this example, 5 out of the 9 leaf nodes are visited and 9 out of the 15 itemsets are compared against the transaction.
6.2.5
Computational Complexity
The computational complexity of the Apriori algorithm can be aﬀected by the following factors. Support Threshold Lowering the support threshold often results in more itemsets being declared as frequent. This has an adverse eﬀect on the com
345
Chapter 6
Association Analysis 1+ 2356
Transaction
2+
356
12356
3+ 56
Candidate Hash Tree 12+
356
13+ 56
234 567
15+ 6 145
136
345
356
367
357
368
689 124
125
457
458
159
Figure 6.12. Subset operation on the leftmost subtree of the root of a candidate hash tree.
putational complexity of the algorithm because more candidate itemsets must be generated and counted, as shown in Figure 6.13. The maximum size of frequent itemsets also tends to increase with lower support thresholds. As the maximum size of the frequent itemsets increases, the algorithm will need to make more passes over the data set. Number of Items (Dimensionality) As the number of items increases, more space will be needed to store the support counts of items. If the number of frequent items also grows with the dimensionality of the data, the computation and I/O costs will increase because of the larger number of candidate itemsets generated by the algorithm. Number of Transactions Since the Apriori algorithm makes repeated passes over the data set, its run time increases with a larger number of transactions. Average Transaction Width For dense data sets, the average transaction width can be very large. This aﬀects the complexity of the Apriori algorithm in two ways. First, the maximum size of frequent itemsets tends to increase as the
346
6.2
Frequent Itemset Generation
×105
4
Support = 0.1% Support = 0.2% Support = 0.5%
Number of Candidate Itemsets
3.5 3
2.5 2
1.5 1
0.5 0 0
5
10 Size of Itemset
15
20
(a) Number of candidate itemsets.
×105
4
Support = 0.1% Support = 0.2% Support = 0.5%
Number of Frequent Itemsets
3.5 3 2.5 2 1.5 1 0.5 0
0
5
10 Size of Itemset
15
20
(b) Number of frequent itemsets.
Figure 6.13. Effect of support threshold on the number of candidate and frequent itemsets.
average transaction width increases. As a result, more candidate itemsets must be examined during candidate generation and support counting, as illustrated in Figure 6.14. Second, as the transaction width increases, more itemsets
347
Chapter 6
Association Analysis ×105
10
Width = 5 Width = 10 Width = 15
9
Number of Candidate Itemsets
8 7 6 5 4 3 2 1 0 0
5
10
15
20
25
Size of Itemset
(a) Number of candidate itemsets.
10
×105 Width = 5 Width = 10 Width = 15
9
Number of Frequent Itemsets
8 7 6 5 4 3 2 1 0 0
5
10 15 Size of Itemset
20
25
(b) Number of Frequent Itemsets.
Figure 6.14. Effect of average transaction width on the number of candidate and frequent itemsets.
are contained in the transaction. This will increase the number of hash tree traversals performed during support counting. A detailed analysis of the time complexity for the Apriori algorithm is presented next.
348
6.3
Rule Generation
Generation of frequent 1itemsets For each transaction, we need to update the support count for every item present in the transaction. Assuming that w is the average transaction width, this operation requires O(N w) time, where N is the total number of transactions. Candidate generation To generate candidate kitemsets, pairs of frequent (k − 1)itemsets are merged to determine whether they have at least k − 2 items in common. Each merging operation requires at most k − 2 equality comparisons. In the bestcase scenario, every merging step produces a viable candidate kitemset. In the worstcase scenario, the algorithm must merge every pair of frequent (k −1)itemsets found in the previous iteration. Therefore, the overall cost of merging frequent itemsets is w
(k − 2)Ck  < Cost of merging
{ }
bcd=>a
cd=>ab
bd=>ac
d=>abc
acd=>b
bc=>ad
c=>abd
abd=>c
ad=>bc
abc=>d
ac=>bd
b=>acd
ab=>cd
a=>bcd
Pruned Rules
Figure 6.15. Pruning of association rules using the confidence measure.
entire subgraph spanned by the node can be pruned immediately. Suppose the conﬁdence for {bcd} −→ {a} is low. All the rules containing item a in its consequent, including {cd} −→ {ab}, {bd} −→ {ac}, {bc} −→ {ad}, and {d} −→ {abc} can be discarded. A pseudocode for the rule generation step is shown in Algorithms 6.2 and 6.3. Note the similarity between the apgenrules procedure given in Algorithm 6.3 and the frequent itemset generation procedure given in Algorithm 6.1. The only diﬀerence is that, in rule generation, we do not have to make additional passes over the data set to compute the conﬁdence of the candidate rules. Instead, we determine the conﬁdence of each rule by using the support counts computed during frequent itemset generation. Algorithm 6.2 Rule generation of the Apriori algorithm. 1: for each frequent kitemset fk , k ≥ 2 do 2: H1 = {i  i ∈ fk } {1item consequents of the rule.} 3: call apgenrules(fk , H1 .) 4: end for
351
Chapter 6
Association Analysis
Algorithm 6.3 Procedure apgenrules(fk , Hm ). 1: 2: 3: 4: 5: 6: 7: 8: 9: 10: 11: 12: 13: 14:
k = fk  {size of frequent itemset.} m = Hm  {size of rule consequent.} if k > m + 1 then Hm+1 = apriorigen(Hm ). for each hm+1 ∈ Hm+1 do conf = σ(fk )/σ(fk − hm+1 ). if conf ≥ minconf then output the rule (fk − hm+1 ) −→ hm+1 . else delete hm+1 from Hm+1 . end if end for call apgenrules(fk , Hm+1 .) end if
6.3.3
An Example: Congressional Voting Records
This section demonstrates the results of applying association analysis to the voting records of members of the United States House of Representatives. The data is obtained from the 1984 Congressional Voting Records Database, which is available at the UCI machine learning data repository. Each transaction contains information about the party aﬃliation for a representative along with his or her voting record on 16 key issues. There are 435 transactions and 34 items in the data set. The set of items are listed in Table 6.3. The Apriori algorithm is then applied to the data set with minsup = 30% and minconf = 90%. Some of the highconﬁdence rules extracted by the algorithm are shown in Table 6.4. The ﬁrst two rules suggest that most of the members who voted yes for aid to El Salvador and no for budget resolution and MX missile are Republicans; while those who voted no for aid to El Salvador and yes for budget resolution and MX missile are Democrats. These highconﬁdence rules show the key issues that divide members from both political parties. If minconf is reduced, we may ﬁnd rules that contain issues that cut across the party lines. For example, with minconf = 40%, the rules suggest that corporation cutbacks is an issue that receives almost equal number of votes from both parties—52.3% of the members who voted no are Republicans, while the remaining 47.7% of them who voted no are Democrats.
352
6.4
Compact Representation of Frequent Itemsets
Table 6.3. List of binary attributes from the 1984 United States Congressional Voting Records. Source: The UCI machine learning repository. 1. Republican 2. Democrat 3. handicappedinfants = yes 4. handicappedinfants = no 5. water project cost sharing = yes 6. water project cost sharing = no 7. budgetresolution = yes 8. budgetresolution = no 9. physician fee freeze = yes 10. physician fee freeze = no 11. aid to El Salvador = yes 12. aid to El Salvador = no 13. religious groups in schools = yes 14. religious groups in schools = no 15. antisatellite test ban = yes 16. antisatellite test ban = no 17. aid to Nicaragua = yes
18. 19. 20. 21. 22. 23. 24. 25. 26. 27. 28. 29. 30. 31. 32. 33. 34.
aid to Nicaragua = no MXmissile = yes MXmissile = no immigration = yes immigration = no synfuel corporation cutback = yes synfuel corporation cutback = no education spending = yes education spending = no righttosue = yes righttosue = no crime = yes crime = no dutyfreeexports = yes dutyfreeexports = no export administration act = yes export administration act = no
Table 6.4. Association rules extracted from the 1984 United States Congressional Voting Records. Association Rule {budget resolution = no, MXmissile=no, aid to El Salvador = yes } −→ {Republican} {budget resolution = yes, MXmissile=yes, aid to El Salvador = no } −→ {Democrat} {crime = yes, righttosue = yes, physician fee freeze = yes} −→ {Republican} {crime = no, righttosue = no, physician fee freeze = no} −→ {Democrat}
6.4
Conﬁdence 91.0% 97.5% 93.5% 100%
Compact Representation of Frequent Itemsets
In practice, the number of frequent itemsets produced from a transaction data set can be very large. It is useful to identify a small representative set of itemsets from which all other frequent itemsets can be derived. Two such representations are presented in this section in the form of maximal and closed frequent itemsets.
353
Chapter 6
Association Analysis null
Maximal Frequent Itemset a
b
c
d
e
ab
ac
ad
ae
bc
bd
be
cd
ce
de
abc
abd
abe
acd
ace
ade
bcd
bce
bde
cde
abcd
abce
abde
acde
bcde Frequent
abcde Infrequent
Frequent Itemset Border
Figure 6.16. Maximal frequent itemset.
6.4.1
Maximal Frequent Itemsets
Deﬁnition 6.3 (Maximal Frequent Itemset). A maximal frequent itemset is deﬁned as a frequent itemset for which none of its immediate supersets are frequent. To illustrate this concept, consider the itemset lattice shown in Figure 6.16. The itemsets in the lattice are divided into two groups: those that are frequent and those that are infrequent. A frequent itemset border, which is represented by a dashed line, is also illustrated in the diagram. Every itemset located above the border is frequent, while those located below the border (the shaded nodes) are infrequent. Among the itemsets residing near the border, {a, d}, {a, c, e}, and {b, c, d, e} are considered to be maximal frequent itemsets because their immediate supersets are infrequent. An itemset such as {a, d} is maximal frequent because all of its immediate supersets, {a, b, d}, {a, c, d}, and {a, d, e}, are infrequent. In contrast, {a, c} is nonmaximal because one of its immediate supersets, {a, c, e}, is frequent. Maximal frequent itemsets eﬀectively provide a compact representation of frequent itemsets. In other words, they form the smallest set of itemsets from
354
6.4
Compact Representation of Frequent Itemsets
which all frequent itemsets can be derived. For example, the frequent itemsets shown in Figure 6.16 can be divided into two groups: • Frequent itemsets that begin with item a and that may contain items c, d, or e. This group includes itemsets such as {a}, {a, c}, {a, d}, {a, e}, and {a, c, e}. • Frequent itemsets that begin with items b, c, d, or e. This group includes itemsets such as {b}, {b, c}, {c, d},{b, c, d, e}, etc. Frequent itemsets that belong in the ﬁrst group are subsets of either {a, c, e} or {a, d}, while those that belong in the second group are subsets of {b, c, d, e}. Hence, the maximal frequent itemsets {a, c, e}, {a, d}, and {b, c, d, e} provide a compact representation of the frequent itemsets shown in Figure 6.16. Maximal frequent itemsets provide a valuable representation for data sets that can produce very long, frequent itemsets, as there are exponentially many frequent itemsets in such data. Nevertheless, this approach is practical only if an eﬃcient algorithm exists to explicitly ﬁnd the maximal frequent itemsets without having to enumerate all their subsets. We brieﬂy describe one such approach in Section 6.5. Despite providing a compact representation, maximal frequent itemsets do not contain the support information of their subsets. For example, the support of the maximal frequent itemsets {a, c, e}, {a, d}, and {b,c,d,e} do not provide any hint about the support of their subsets. An additional pass over the data set is therefore needed to determine the support counts of the nonmaximal frequent itemsets. In some cases, it might be desirable to have a minimal representation of frequent itemsets that preserves the support information. We illustrate such a representation in the next section.
6.4.2
Closed Frequent Itemsets
Closed itemsets provide a minimal representation of itemsets without losing their support information. A formal deﬁnition of a closed itemset is presented below. Deﬁnition 6.4 (Closed Itemset). An itemset X is closed if none of its immediate supersets has exactly the same support count as X. Put another way, X is not closed if at least one of its immediate supersets has the same support count as X. Examples of closed itemsets are shown in Figure 6.17. To better illustrate the support count of each itemset, we have associated each node (itemset) in the lattice with a list of its corresponding
355
Chapter 6 TID
Items
1
abc
2
abcd
3
bce
4
acde
5
de
1,2 ab
1,2 abc
Association Analysis minsup = 40% null 1,2,4 a
1,2,4 ac
2,4 ad
2
1,2,3 b
1,2,3
4 ae
2,4 abd
1,2,3,4 c
abe
2 bc
3
bd
4 acd
2,4,5 d
4 ace
2
2,4 be
2 ade
3,4,5 e
cd
3,4
ce
3 bcd
4,5
de
4 bce
bde
cde
4 abcd
Closed Frequent Itemset
abce
abde
acde
bcde
abcde
Figure 6.17. An example of the closed frequent itemsets (with minimum support count equal to 40%).
transaction IDs. For example, since the node {b, c} is associated with transaction IDs 1, 2, and 3, its support count is equal to three. From the transactions given in this diagram, notice that every transaction that contains b also contains c. Consequently, the support for {b} is identical to {b, c} and {b} should not be considered a closed itemset. Similarly, since c occurs in every transaction that contains both a and d, the itemset {a, d} is not closed. On the other hand, {b, c} is a closed itemset because it does not have the same support count as any of its supersets. Deﬁnition 6.5 (Closed Frequent Itemset). An itemset is a closed frequent itemset if it is closed and its support is greater than or equal to minsup. In the previous example, assuming that the support threshold is 40%, {b,c} is a closed frequent itemset because its support is 60%. The rest of the closed frequent itemsets are indicated by the shaded nodes. Algorithms are available to explicitly extract closed frequent itemsets from a given data set. Interested readers may refer to the bibliographic notes at the end of this chapter for further discussions of these algorithms. We can use the closed frequent itemsets to determine the support counts for the nonclosed
356
6.4
Compact Representation of Frequent Itemsets
Algorithm 6.4 Support counting using closed frequent itemsets. 1: 2: 3: 4: 5: 6: 7: 8: 9: 10: 11:
Let C denote the set of closed frequent itemsets Let kmax denote the maximum size of closed frequent itemsets Fkmax = {f f ∈ C, f  = kmax } {Find all frequent itemsets of size kmax .} for k = kmax − 1 downto 1 do Fk = {f f ⊂ Fk+1 , f  = k} {Find all frequent itemsets of size k.} for each f ∈ Fk do if f ∈ / C then f.support = max{f .supportf ∈ Fk+1 , f ⊂ f } end if end for end for
frequent itemsets. For example, consider the frequent itemset {a, d} shown in Figure 6.17. Because the itemset is not closed, its support count must be identical to one of its immediate supersets. The key is to determine which superset (among {a, b, d}, {a, c, d}, or {a, d, e}) has exactly the same support count as {a, d}. The Apriori principle states that any transaction that contains the superset of {a, d} must also contain {a, d}. However, any transaction that contains {a, d} does not have to contain the supersets of {a, d}. For this reason, the support for {a, d} must be equal to the largest support among its supersets. Since {a, c, d} has a larger support than both {a, b, d} and {a, d, e}, the support for {a, d} must be identical to the support for {a, c, d}. Using this methodology, an algorithm can be developed to compute the support for the nonclosed frequent itemsets. The pseudocode for this algorithm is shown in Algorithm 6.4. The algorithm proceeds in a speciﬁctogeneral fashion, i.e., from the largest to the smallest frequent itemsets. This is because, in order to ﬁnd the support for a nonclosed frequent itemset, the support for all of its supersets must be known. To illustrate the advantage of using closed frequent itemsets, consider the data set shown in Table 6.5, which contains ten transactions and ﬁfteen items. The items can be divided into three groups: (1) Group A, which contains items a1 through a5 ; (2) Group B, which contains items b1 through b5 ; and (3) Group C, which contains items c1 through c5 . Note that items within each group are perfectly associated with each other and they do not appear with items from another group. Assuming the support threshold is 20%, the total number of frequent itemsets is 3 × (25 − 1) = 93. However, there are only three closed frequent itemsets in the data: ({a1 , a2 , a3 , a4 , a5 }, {b1 , b2 , b3 , b4 , b5 }, and {c1 , c2 , c3 , c4 , c5 }). It is often suﬃcient to present only the closed frequent itemsets to the analysts instead of the entire set of frequent itemsets.
357
Chapter 6
Association Analysis
Table 6.5. A transaction data set for mining closed itemsets. TID 1 2 3 4 5 6 7 8 9 10
a1 1 1 1 0 0 0 0 0 0 0
a2 1 1 1 0 0 0 0 0 0 0
a3 1 1 1 0 0 0 0 0 0 0
a4 1 1 1 0 0 0 0 0 0 0
a5 1 1 1 0 0 0 0 0 0 0
b1 0 0 0 1 1 1 0 0 0 0
b2 0 0 0 1 1 1 0 0 0 0
b3 0 0 0 1 1 1 0 0 0 0
b4 0 0 0 1 1 1 0 0 0 0
b5 0 0 0 1 1 1 0 0 0 0
c1 0 0 0 0 0 0 1 1 1 1
c2 0 0 0 0 0 0 1 1 1 1
c3 0 0 0 0 0 0 1 1 1 1
c4 0 0 0 0 0 0 1 1 1 1
c5 0 0 0 0 0 0 1 1 1 1
Frequent Itemsets Closed Frequent Itemsets Maximal Frequent Itemsets
Figure 6.18. Relationships among frequent, maximal frequent, and closed frequent itemsets.
Closed frequent itemsets are useful for removing some of the redundant association rules. An association rule X −→ Y is redundant if there exists another rule X −→ Y , where X is a subset of X and Y is a subset of Y , such that the support and conﬁdence for both rules are identical. In the example shown in Figure 6.17, {b} is not a closed frequent itemset while {b, c} is closed. The association rule {b} −→ {d, e} is therefore redundant because it has the same support and conﬁdence as {b, c} −→ {d, e}. Such redundant rules are not generated if closed frequent itemsets are used for rule generation. Finally, note that all maximal frequent itemsets are closed because none of the maximal frequent itemsets can have the same support count as their immediate supersets. The relationships among frequent, maximal frequent, and closed frequent itemsets are shown in Figure 6.18.
358
6.5
6.5
Alternative Methods for Generating Frequent Itemsets
Alternative Methods for Generating Frequent Itemsets
Apriori is one of the earliest algorithms to have successfully addressed the combinatorial explosion of frequent itemset generation. It achieves this by applying the Apriori principle to prune the exponential search space. Despite its signiﬁcant performance improvement, the algorithm still incurs considerable I/O overhead since it requires making several passes over the transaction data set. In addition, as noted in Section 6.2.5, the performance of the Apriori algorithm may degrade signiﬁcantly for dense data sets because of the increasing width of transactions. Several alternative methods have been developed to overcome these limitations and improve upon the eﬃciency of the Apriori algorithm. The following is a highlevel description of these methods. Traversal of Itemset Lattice A search for frequent itemsets can be conceptually viewed as a traversal on the itemset lattice shown in Figure 6.1. The search strategy employed by an algorithm dictates how the lattice structure is traversed during the frequent itemset generation process. Some search strategies are better than others, depending on the conﬁguration of frequent itemsets in the lattice. An overview of these strategies is presented next. • GeneraltoSpeciﬁc versus SpeciﬁctoGeneral: The Apriori algorithm uses a generaltospeciﬁc search strategy, where pairs of frequent (k−1)itemsets are merged to obtain candidate kitemsets. This generaltospeciﬁc search strategy is eﬀective, provided the maximum length of a frequent itemset is not too long. The conﬁguration of frequent itemsets that works best with this strategy is shown in Figure 6.19(a), where the darker nodes represent infrequent itemsets. Alternatively, a speciﬁctogeneral search strategy looks for more speciﬁc frequent itemsets ﬁrst, before ﬁnding the more general frequent itemsets. This strategy is useful to discover maximal frequent itemsets in dense transactions, where the frequent itemset border is located near the bottom of the lattice, as shown in Figure 6.19(b). The Apriori principle can be applied to prune all subsets of maximal frequent itemsets. Speciﬁcally, if a candidate kitemset is maximal frequent, we do not have to examine any of its subsets of size k − 1. However, if the candidate kitemset is infrequent, we need to check all of its k − 1 subsets in the next iteration. Another approach is to combine both generaltospeciﬁc and speciﬁctogeneral search strategies. This bidirectional approach requires more space to
359
Chapter 6
Association Analysis
Frequent Itemset Border null
{a1,a2,...,an} (a) Generaltospecific
Frequent Itemset Border
null
{a1,a2,...,an}
Frequent Itemset Border
(b) Specifictogeneral
null
{a1,a2,...,an} (c) Bidirectional
Figure 6.19. Generaltospecific, specifictogeneral, and bidirectional search.
store the candidate itemsets, but it can help to rapidly identify the frequent itemset border, given the conﬁguration shown in Figure 6.19(c). • Equivalence Classes: Another way to envision the traversal is to ﬁrst partition the lattice into disjoint groups of nodes (or equivalence classes). A frequent itemset generation algorithm searches for frequent itemsets within a particular equivalence class ﬁrst before moving to another equivalence class. As an example, the levelwise strategy used in the Apriori algorithm can be considered to be partitioning the lattice on the basis of itemset sizes; i.e., the algorithm discovers all frequent 1itemsets ﬁrst before proceeding to largersized itemsets. Equivalence classes can also be deﬁned according to the preﬁx or suﬃx labels of an itemset. In this case, two itemsets belong to the same equivalence class if they share a common preﬁx or suﬃx of length k. In the preﬁxbased approach, the algorithm can search for frequent itemsets starting with the preﬁx a before looking for those starting with preﬁxes b, c, and so on. Both preﬁxbased and suﬃxbased equivalence classes can be demonstrated using the treelike structure shown in Figure 6.20. • BreadthFirst versus DepthFirst: The Apriori algorithm traverses the lattice in a breadthﬁrst manner, as shown in Figure 6.21(a). It ﬁrst discovers all the frequent 1itemsets, followed by the frequent 2itemsets, and so on, until no new frequent itemsets are generated. The itemset
360
6.5
Alternative Methods for Generating Frequent Itemsets null
b
a
ab
ac
ad
abc
abd
acd
null
c
bc
bd
bcd
d
cd
a
ab
ac
abc
b
c
bc
ad
abd
d
bd
acd
abcd
cd
bcd
abcd
(a) Prefix tree.
(b) Suffix tree.
Figure 6.20. Equivalence classes based on the prefix and suffix labels of itemsets.
(a) Breadth first
(b) Depth first
Figure 6.21. Breadthfirst and depthfirst traversals.
lattice can also be traversed in a depthﬁrst manner, as shown in Figures 6.21(b) and 6.22. The algorithm can start from, say, node a in Figure 6.22, and count its support to determine whether it is frequent. If so, the algorithm progressively expands the next level of nodes, i.e., ab, abc, and so on, until an infrequent node is reached, say, abcd. It then backtracks to another branch, say, abce, and continues the search from there. The depthﬁrst approach is often used by algorithms designed to ﬁnd maximal frequent itemsets. This approach allows the frequent itemset border to be detected more quickly than using a breadthﬁrst approach. Once a maximal frequent itemset is found, substantial pruning can be
361
Chapter 6
Association Analysis null
a
b
c
d e
ab
ac
ad ae
abc
abd
bc
cd be
acd abe
bd
ce
de
bcd ace
bce
ade
bde
cde
abcd abce
abde
acde
bcde
abcde
Figure 6.22. Generating candidate itemsets using the depthfirst approach.
performed on its subsets. For example, if the node bcde shown in Figure 6.22 is maximal frequent, then the algorithm does not have to visit the subtrees rooted at bd, be, c, d, and e because they will not contain any maximal frequent itemsets. However, if abc is maximal frequent, only the nodes such as ac and bc are not maximal frequent (but the subtrees of ac and bc may still contain maximal frequent itemsets). The depthﬁrst approach also allows a diﬀerent kind of pruning based on the support of itemsets. For example, suppose the support for {a, b, c} is identical to the support for {a, b}. The subtrees rooted at abd and abe can be skipped because they are guaranteed not to have any maximal frequent itemsets. The proof of this is left as an exercise to the readers. Representation of Transaction Data Set There are many ways to represent a transaction data set. The choice of representation can aﬀect the I/O costs incurred when computing the support of candidate itemsets. Figure 6.23 shows two diﬀerent ways of representing market basket transactions. The representation on the left is called a horizontal data layout, which is adopted by many association rule mining algorithms, including Apriori. Another possibility is to store the list of transaction identiﬁers (TIDlist) associated with each item. Such a representation is known as the vertical data layout. The support for each candidate itemset is obtained by intersecting the TIDlists of its subset items. The length of the TIDlists shrinks as we progress to larger
362
6.6 Horizontal Data Layout TID 1 2 3 4 5 6 7 8 9 10
Items a,b,e b,c,d c,e a,c,d a,b,c,d a,e a,b a,b,c a,c,d b
FPGrowth Algorithm
Vertical Data Layout a 1 4 5 6 7 8 9
b 1 2 5 7 8 10
c 2 3 4 8 9
d 2 4 5 9
e 1 3 6
Figure 6.23. Horizontal and vertical data format.
sized itemsets. However, one problem with this approach is that the initial set of TIDlists may be too large to ﬁt into main memory, thus requiring more sophisticated techniques to compress the TIDlists. We describe another eﬀective approach to represent the data in the next section.
6.6
FPGrowth Algorithm
This section presents an alternative algorithm called FPgrowth that takes a radically diﬀerent approach to discovering frequent itemsets. The algorithm does not subscribe to the generateandtest paradigm of Apriori. Instead, it encodes the data set using a compact data structure called an FPtree and extracts frequent itemsets directly from this structure. The details of this approach are presented next.
6.6.1
FPTree Representation
An FPtree is a compressed representation of the input data. It is constructed by reading the data set one transaction at a time and mapping each transaction onto a path in the FPtree. As diﬀerent transactions can have several items in common, their paths may overlap. The more the paths overlap with one another, the more compression we can achieve using the FPtree structure. If the size of the FPtree is small enough to ﬁt into main memory, this will allow us to extract frequent itemsets directly from the structure in memory instead of making repeated passes over the data stored on disk.
363
Chapter 6
Association Analysis
Transaction Data Set TID Items {a,b} 1 2 {b,c,d} {a,c,d,e} 3 4 {a,d,e} {a,b,c} 5 {a,b,c,d} 6 7 {a} 8 {a,b,c} 9 {a,b,d} 10 {b,c,e}
null null
a:1
b:1
a:1 c:1
b:1 b:1
d:1 (i) After reading TID=1 (ii) After reading TID=2 null a:2
b:1
b:1
c:1
c:1 d:1
d:1
e:1 (iii) After reading TID=3 null a:8
b:2
b:5
c:2 c:1
c:3 d:1 d:1
d:1 d:1
d:1 e:1
e:1
e:1
(iv) After reading TID=10
Figure 6.24. Construction of an FPtree.
Figure 6.24 shows a data set that contains ten transactions and ﬁve items. The structures of the FPtree after reading the ﬁrst three transactions are also depicted in the diagram. Each node in the tree contains the label of an item along with a counter that shows the number of transactions mapped onto the given path. Initially, the FPtree contains only the root node represented by the null symbol. The FPtree is subsequently extended in the following way: 1. The data set is scanned once to determine the support count of each item. Infrequent items are discarded, while the frequent items are sorted in decreasing support counts. For the data set shown in Figure 6.24, a is the most frequent item, followed by b, c, d, and e.
364
6.6
FPGrowth Algorithm
2. The algorithm makes a second pass over the data to construct the FPtree. After reading the ﬁrst transaction, {a, b}, the nodes labeled as a and b are created. A path is then formed from null → a → b to encode the transaction. Every node along the path has a frequency count of 1. 3. After reading the second transaction, {b,c,d}, a new set of nodes is created for items b, c, and d. A path is then formed to represent the transaction by connecting the nodes null → b → c → d. Every node along this path also has a frequency count equal to one. Although the ﬁrst two transactions have an item in common, which is b, their paths are disjoint because the transactions do not share a common preﬁx. 4. The third transaction, {a,c,d,e}, shares a common preﬁx item (which is a) with the ﬁrst transaction. As a result, the path for the third transaction, null → a → c → d → e, overlaps with the path for the ﬁrst transaction, null → a → b. Because of their overlapping path, the frequency count for node a is incremented to two, while the frequency counts for the newly created nodes, c, d, and e, are equal to one. 5. This process continues until every transaction has been mapped onto one of the paths given in the FPtree. The resulting FPtree after reading all the transactions is shown at the bottom of Figure 6.24. The size of an FPtree is typically smaller than the size of the uncompressed data because many transactions in market basket data often share a few items in common. In the bestcase scenario, where all the transactions have the same set of items, the FPtree contains only a single branch of nodes. The worstcase scenario happens when every transaction has a unique set of items. As none of the transactions have any items in common, the size of the FPtree is eﬀectively the same as the size of the original data. However, the physical storage requirement for the FPtree is higher because it requires additional space to store pointers between nodes and counters for each item. The size of an FPtree also depends on how the items are ordered. If the ordering scheme in the preceding example is reversed, i.e., from lowest to highest support item, the resulting FPtree is shown in Figure 6.25. The tree appears to be denser because the branching factor at the root node has increased from 2 to 5 and the number of nodes containing the high support items such as a and b has increased from 3 to 12. Nevertheless, ordering by decreasing support counts does not always lead to the smallest tree. For example, suppose we augment the data set given in Figure 6.24 with 100 transactions that contain {e}, 80 transactions that contain {d}, 60 transactions
365
Chapter 6
Association Analysis null
d:3
e:3
b:1
c:2
d:2 c:1
c:1
a:1
a:1 b:1
b:2
b:2
b:1 a:1
c:2
a:1 a:1
a:2
a:1
Figure 6.25. An FPtree representation for the data set shown in Figure 6.24 with a different item ordering scheme.
that contain {c}, and 40 transactions that contain {b}. Item e is now most frequent, followed by d, c, b, and a. With the augmented transactions, ordering by decreasing support counts will result in an FPtree similar to Figure 6.25, while a scheme based on increasing support counts produces a smaller FPtree similar to Figure 6.24(iv). An FPtree also contains a list of pointers connecting between nodes that have the same items. These pointers, represented as dashed lines in Figures 6.24 and 6.25, help to facilitate the rapid access of individual items in the tree. We explain how to use the FPtree and its corresponding pointers for frequent itemset generation in the next section.
6.6.2
Frequent Itemset Generation in FPGrowth Algorithm
FPgrowth is an algorithm that generates frequent itemsets from an FPtree by exploring the tree in a bottomup fashion. Given the example tree shown in Figure 6.24, the algorithm looks for frequent itemsets ending in e ﬁrst, followed by d, c, b, and ﬁnally, a. This bottomup strategy for ﬁnding frequent itemsets ending with a particular item is equivalent to the suﬃxbased approach described in Section 6.5. Since every transaction is mapped onto a path in the FPtree, we can derive the frequent itemsets ending with a particular item, say, e, by examining only the paths containing node e. These paths can be accessed rapidly using the pointers associated with node e. The extracted paths are shown in Figure 6.26(a). The details on how to process the paths to obtain frequent itemsets will be explained later.
366
6.6
FPGrowth Algorithm null
null a:8 a:8
b:2 b:5
c:1
b:2
c:2 d:1
c:3
c:2 c:1
d:1 d:1
e:1 e:1 e:1 (a) Paths containing node e
null
b:2 c:2 c:1
d:1
d:1
null
a:8 b:2
b:5
d:1
(b) Paths containing node d
null
a:8
d:1
a:8
b:5
c:3 (c) Paths containing node c
(d) Paths containing node b
(e) Paths containing node a
Figure 6.26. Decomposing the frequent itemset generation problem into multiple subproblems, where each subproblem involves finding frequent itemsets ending in e, d, c, b, and a.
Table 6.6. The list of frequent itemsets ordered by their corresponding suffixes.
Suﬃx e d c b a
Frequent Itemsets {e}, {d,e}, {a,d,e}, {c,e},{a,e} {d}, {c,d}, {b,c,d}, {a,c,d}, {b,d}, {a,b,d}, {a,d} {c}, {b,c}, {a,b,c}, {a,c} {b}, {a,b} {a}
After ﬁnding the frequent itemsets ending in e, the algorithm proceeds to look for frequent itemsets ending in d by processing the paths associated with node d. The corresponding paths are shown in Figure 6.26(b). This process continues until all the paths associated with nodes c, b, and ﬁnally a, are processed. The paths for these items are shown in Figures 6.26(c), (d), and (e), while their corresponding frequent itemsets are summarized in Table 6.6. FPgrowth ﬁnds all the frequent itemsets ending with a particular suﬃx by employing a divideandconquer strategy to split the problem into smaller subproblems. For example, suppose we are interested in ﬁnding all frequent
367
Chapter 6
Association Analysis null
null
a:8
b:2
c:1
a:2
c:2 c:1
d:1
c:1
d:1 e:1
e:1
e:1
(a) Prefix paths ending in e null
d:1
d:1
(b) Conditional FPtree for e null
a:2 c:1 a:2
d:1
d:1
(c) Prefix paths ending in de
(d) Conditional FPtree for de null
null a:2 c:1
c:1
(e) Prefix paths ending in ce
a:2 (f) Prefix paths ending in ae
Figure 6.27. Example of applying the FPgrowth algorithm to find frequent itemsets ending in e.
itemsets ending in e. To do this, we must ﬁrst check whether the itemset {e} itself is frequent. If it is frequent, we consider the subproblem of ﬁnding frequent itemsets ending in de, followed by ce, be, and ae. In turn, each of these subproblems are further decomposed into smaller subproblems. By merging the solutions obtained from the subproblems, all the frequent itemsets ending in e can be found. This divideandconquer approach is the key strategy employed by the FPgrowth algorithm. For a more concrete example on how to solve the subproblems, consider the task of ﬁnding frequent itemsets ending with e. 1. The ﬁrst step is to gather all the paths containing node e. These initial paths are called preﬁx paths and are shown in Figure 6.27(a). 2. From the preﬁx paths shown in Figure 6.27(a), the support count for e is obtained by adding the support counts associated with node e. Assuming that the minimum support count is 2, {e} is declared a frequent itemset because its support count is 3.
368
6.6
FPGrowth Algorithm
3. Because {e} is frequent, the algorithm has to solve the subproblems of ﬁnding frequent itemsets ending in de, ce, be, and ae. Before solving these subproblems, it must ﬁrst convert the preﬁx paths into a conditional FPtree, which is structurally similar to an FPtree, except it is used to ﬁnd frequent itemsets ending with a particular suﬃx. A conditional FPtree is obtained in the following way: (a) First, the support counts along the preﬁx paths must be updated because some of the counts include transactions that do not contain item e. For example, the rightmost path shown in Figure 6.27(a), null −→ b:2 −→ c:2 −→ e:1, includes a transaction {b, c} that does not contain item e. The counts along the preﬁx path must therefore be adjusted to 1 to reﬂect the actual number of transactions containing {b, c, e}. (b) The preﬁx paths are truncated by removing the nodes for e. These nodes can be removed because the support counts along the preﬁx paths have been updated to reﬂect only transactions that contain e and the subproblems of ﬁnding frequent itemsets ending in de, ce, be, and ae no longer need information about node e. (c) After updating the support counts along the preﬁx paths, some of the items may no longer be frequent. For example, the node b appears only once and has a support count equal to 1, which means that there is only one transaction that contains both b and e. Item b can be safely ignored from subsequent analysis because all itemsets ending in be must be infrequent. The conditional FPtree for e is shown in Figure 6.27(b). The tree looks diﬀerent than the original preﬁx paths because the frequency counts have been updated and the nodes b and e have been eliminated. 4. FPgrowth uses the conditional FPtree for e to solve the subproblems of ﬁnding frequent itemsets ending in de, ce, and ae. To ﬁnd the frequent itemsets ending in de, the preﬁx paths for d are gathered from the conditional FPtree for e (Figure 6.27(c)). By adding the frequency counts associated with node d, we obtain the support count for {d, e}. Since the support count is equal to 2, {d, e} is declared a frequent itemset. Next, the algorithm constructs the conditional FPtree for de using the approach described in step 3. After updating the support counts and removing the infrequent item c, the conditional FPtree for de is shown in Figure 6.27(d). Since the conditional FPtree contains only one item,
369
Chapter 6
Association Analysis
a, whose support is equal to minsup, the algorithm extracts the frequent itemset {a, d, e} and moves on to the next subproblem, which is to generate frequent itemsets ending in ce. After processing the preﬁx paths for c, only {c, e} is found to be frequent. The algorithm proceeds to solve the next subprogram and found {a, e} to be the only frequent itemset remaining. This example illustrates the divideandconquer approach used in the FPgrowth algorithm. At each recursive step, a conditional FPtree is constructed by updating the frequency counts along the preﬁx paths and removing all infrequent items. Because the subproblems are disjoint, FPgrowth will not generate any duplicate itemsets. In addition, the counts associated with the nodes allow the algorithm to perform support counting while generating the common suﬃx itemsets. FPgrowth is an interesting algorithm because it illustrates how a compact representation of the transaction data set helps to eﬃciently generate frequent itemsets. In addition, for certain transaction data sets, FPgrowth outperforms the standard Apriori algorithm by several orders of magnitude. The runtime performance of FPgrowth depends on the compaction factor of the data set. If the resulting conditional FPtrees are very bushy (in the worst case, a full preﬁx tree), then the performance of the algorithm degrades signiﬁcantly because it has to generate a large number of subproblems and merge the results returned by each subproblem.
6.7
Evaluation of Association Patterns
Association analysis algorithms have the potential to generate a large number of patterns. For example, although the data set shown in Table 6.1 contains only six items, it can produce up to hundreds of association rules at certain support and conﬁdence thresholds. As the size and dimensionality of real commercial databases can be very large, we could easily end up with thousands or even millions of patterns, many of which might not be interesting. Sifting through the patterns to identify the most interesting ones is not a trivial task because “one person’s trash might be another person’s treasure.” It is therefore important to establish a set of wellaccepted criteria for evaluating the quality of association patterns. The ﬁrst set of criteria can be established through statistical arguments. Patterns that involve a set of mutually independent items or cover very few transactions are considered uninteresting because they may capture spurious relationships in the data. Such patterns can be eliminated by applying an
370
6.7
Evaluation of Association Patterns
objective interestingness measure that uses statistics derived from data to determine whether a pattern is interesting. Examples of objective interestingness measures include support, conﬁdence, and correlation. The second set of criteria can be established through subjective arguments. A pattern is considered subjectively uninteresting unless it reveals unexpected information about the data or provides useful knowledge that can lead to proﬁtable actions. For example, the rule {Butter} −→ {Bread} may not be interesting, despite having high support and conﬁdence values, because the relationship represented by the rule may seem rather obvious. On the other hand, the rule {Diapers} −→ {Beer} is interesting because the relationship is quite unexpected and may suggest a new crossselling opportunity for retailers. Incorporating subjective knowledge into pattern evaluation is a diﬃcult task because it requires a considerable amount of prior information from the domain experts. The following are some of the approaches for incorporating subjective knowledge into the pattern discovery task. Visualization This approach requires a userfriendly environment to keep the human user in the loop. It also allows the domain experts to interact with the data mining system by interpreting and verifying the discovered patterns. Templatebased approach This approach allows the users to constrain the type of patterns extracted by the mining algorithm. Instead of reporting all the extracted rules, only rules that satisfy a userspeciﬁed template are returned to the users. Subjective interestingness measure A subjective measure can be deﬁned based on domain information such as concept hierarchy (to be discussed in Section 7.3) or proﬁt margin of items. The measure can then be used to ﬁlter patterns that are obvious and nonactionable. Readers interested in subjective interestingness measures may refer to resources listed in the bibliography at the end of this chapter.
6.7.1
Objective Measures of Interestingness
An objective measure is a datadriven approach for evaluating the quality of association patterns. It is domainindependent and requires minimal input from the users, other than to specify a threshold for ﬁltering lowquality patterns. An objective measure is usually computed based on the frequency
371
Chapter 6
Association Analysis
Table 6.7. A 2way contingency table for variables A and B. B
B
A
f11
f10
f1+
A
f01
f00
f0+
f+1
f+0
N
counts tabulated in a contingency table. Table 6.7 shows an example of a contingency table for a pair of binary variables, A and B. We use the notation A (B) to indicate that A (B) is absent from a transaction. Each entry fij in this 2 × 2 table denotes a frequency count. For example, f11 is the number of times A and B appear together in the same transaction, while f01 is the number of transactions that contain B but not A. The row sum f1+ represents the support count for A, while the column sum f+1 represents the support count for B. Finally, even though our discussion focuses mainly on asymmetric binary variables, note that contingency tables are also applicable to other attribute types such as symmetric binary, nominal, and ordinal variables. Limitations of the SupportConﬁdence Framework Existing association rule mining formulation relies on the support and conﬁdence measures to eliminate uninteresting patterns. The drawback of support was previously described in Section 6.8, in which many potentially interesting patterns involving low support items might be eliminated by the support threshold. The drawback of conﬁdence is more subtle and is best demonstrated with the following example. Example 6.3. Suppose we are interested in analyzing the relationship between people who drink tea and coﬀee. We may gather information about the beverage preferences among a group of people and summarize their responses into a table such as the one shown in Table 6.8. Table 6.8. Beverage preferences among a group of 1000 people.
372
Cof f ee
Cof f ee
T ea
150
50
200
T ea
650
150
800
800
200
1000
6.7
Evaluation of Association Patterns
The information given in this table can be used to evaluate the association rule {T ea} −→ {Cof f ee}. At ﬁrst glance, it may appear that people who drink tea also tend to drink coﬀee because the rule’s support (15%) and conﬁdence (75%) values are reasonably high. This argument would have been acceptable except that the fraction of people who drink coﬀee, regardless of whether they drink tea, is 80%, while the fraction of tea drinkers who drink coﬀee is only 75%. Thus knowing that a person is a tea drinker actually decreases her probability of being a coﬀee drinker from 80% to 75%! The rule {T ea} −→ {Cof f ee} is therefore misleading despite its high conﬁdence value. The pitfall of conﬁdence can be traced to the fact that the measure ignores the support of the itemset in the rule consequent. Indeed, if the support of coﬀee drinkers is taken into account, we would not be surprised to ﬁnd that many of the people who drink tea also drink coﬀee. What is more surprising is that the fraction of tea drinkers who drink coﬀee is actually less than the overall fraction of people who drink coﬀee, which points to an inverse relationship between tea drinkers and coﬀee drinkers. Because of the limitations in the supportconﬁdence framework, various objective measures have been used to evaluate the quality of association patterns. Below, we provide a brief description of these measures and explain some of their strengths and limitations. Interest Factor The teacoﬀee example shows that highconﬁdence rules can sometimes be misleading because the conﬁdence measure ignores the support of the itemset appearing in the rule consequent. One way to address this problem is by applying a metric known as lift: Lif t =
c(A −→ B) , s(B)
(6.4)
which computes the ratio between the rule’s conﬁdence and the support of the itemset in the rule consequent. For binary variables, lift is equivalent to another objective measure called interest factor, which is deﬁned as follows: I(A, B) =
N f11 s(A, B) = . s(A) × s(B) f1+ f+1
(6.5)
Interest factor compares the frequency of a pattern against a baseline frequency computed under the statistical independence assumption. The baseline frequency for a pair of mutually independent variables is f1+ f+1 f11 = × , N N N
or equivalently, f11 =
f1+ f+1 . N
(6.6)
373
Chapter 6
Association Analysis
Table 6.9. Contingency tables for the word pairs ({p,q} and {r,s}. p
p
q
880
50
930
q
50
20
70
930
70
1000
r
r
s
20
50
70
s
50
880
930
70
930
1000
This equation follows from the standard approach of using simple fractions as estimates for probabilities. The fraction f11 /N is an estimate for the joint probability P (A, B), while f1+ /N and f+1 /N are the estimates for P (A) and P (B), respectively. If A and B are statistically independent, then P (A, B) = P (A) × P (B), thus leading to the formula shown in Equation 6.6. Using Equations 6.5 and 6.6, we can interpret the measure as follows: = 1, if A and B are independent; > 1, if A and B are positively correlated; I(A, B) < 1, if A and B are negatively correlated.
(6.7)
0.15 = 0.9375, thus sugFor the teacoﬀee example shown in Table 6.8, I = 0.2×0.8 gesting a slight negative correlation between tea drinkers and coﬀee drinkers.
Limitations of Interest Factor We illustrate the limitation of interest factor with an example from the text mining domain. In the text domain, it is reasonable to assume that the association between a pair of words depends on the number of documents that contain both words. For example, because of their stronger association, we expect the words data and mining to appear together more frequently than the words compiler and mining in a collection of computer science articles. Table 6.9 shows the frequency of occurrences between two pairs of words, {p, q} and {r, s}. Using the formula given in Equation 6.5, the interest factor for {p, q} is 1.02 and for {r, s} is 4.08. These results are somewhat troubling for the following reasons. Although p and q appear together in 88% of the documents, their interest factor is close to 1, which is the value when p and q are statistically independent. On the other hand, the interest factor for {r, s} is higher than {p, q} even though r and s seldom appear together in the same document. Conﬁdence is perhaps the better choice in this situation because it considers the association between p and q (94.6%) to be much stronger than that between r and s (28.6%).
374
6.7
Evaluation of Association Patterns
Correlation Analysis Correlation analysis is a statisticalbased technique for analyzing relationships between a pair of variables. For continuous variables, correlation is deﬁned using Pearson’s correlation coeﬃcient (see Equation 2.10 on page 77). For binary variables, correlation can be measured using the φcoeﬃcient, which is deﬁned as f11 f00 − f01 f10 . φ= f1+ f+1 f0+ f+0
(6.8)
The value of correlation ranges from −1 (perfect negative correlation) to +1 (perfect positive correlation). If the variables are statistically independent, then φ = 0. For example, the correlation between the tea and coﬀee drinkers given in Table 6.8 is −0.0625. Limitations of Correlation Analysis The drawback of using correlation can be seen from the word association example given in Table 6.9. Although the words p and q appear together more often than r and s, their φcoeﬃcients are identical, i.e., φ(p, q) = φ(r, s) = 0.232. This is because the φcoeﬃcient gives equal importance to both copresence and coabsence of items in a transaction. It is therefore more suitable for analyzing symmetric binary variables. Another limitation of this measure is that it does not remain invariant when there are proportional changes to the sample size. This issue will be discussed in greater detail when we describe the properties of objective measures on page 377. IS Measure IS is an alternative measure that has been proposed for handling asymmetric binary variables. The measure is deﬁned as follows: IS(A, B) =
s(A, B) I(A, B) × s(A, B) = . s(A)s(B)
(6.9)
Note that IS is large when the interest factor and support of the pattern are large. For example, the value of IS for the word pairs {p, q} and {r, s} shown in Table 6.9 are 0.946 and 0.286, respectively. Contrary to the results given by interest factor and the φcoeﬃcient, the IS measure suggests that the association between {p, q} is stronger than {r, s}, which agrees with what we expect from word associations in documents. It is possible to show that IS is mathematically equivalent to the cosine measure for binary variables (see Equation 2.7 on page 75). In this regard, we
375
Chapter 6
Association Analysis
Table 6.10. Example of a contingency table for items p and q. q
q
p
800
100
900
p
100
0
100
900
100
1000
consider A and B as a pair of bit vectors, A • B = s(A, B) the dot product between the vectors, and A = s(A) the magnitude of vector A. Therefore: s(A, B) A•B IS(A, B) = = = cosine(A, B). A × B s(A) × s(B)
(6.10)
The IS measure can also be expressed as the geometric mean between the conﬁdence of association rules extracted from a pair of binary variables: ! IS(A, B) =
s(A, B) s(A, B) × = c(A → B) × c(B → A). s(A) s(B)
(6.11)
Because the geometric mean between any two numbers is always closer to the smaller number, the IS value of an itemset {p, q} is low whenever one of its rules, p −→ q or q −→ p, has low conﬁdence. Limitations of IS Measure sets, A and B, is
The IS value for a pair of independent item
s(A, B) s(A) × s(B) = = s(A) × s(B). ISindep (A, B) = s(A) × s(B) s(A) × s(B) Since the value depends on s(A) and s(B), IS shares a similar problem as the conﬁdence measure—that the value of the measure can be quite large, even for uncorrelated and negatively correlated patterns. For example, despite the large IS value between items p and q given in Table 6.10 (0.889), it is still less than the expected value when the items are statistically independent (ISindep = 0.9).
376
6.7
Evaluation of Association Patterns
Alternative Objective Interestingness Measures Besides the measures we have described so far, there are other alternative measures proposed for analyzing relationships between pairs of binary variables. These measures can be divided into two categories, symmetric and asymmetric measures. A measure M is symmetric if M (A −→ B) = M (B −→ A). For example, interest factor is a symmetric measure because its value is identical for the rules A −→ B and B −→ A. In contrast, conﬁdence is an asymmetric measure since the conﬁdence for A −→ B and B −→ A may not be the same. Symmetric measures are generally used for evaluating itemsets, while asymmetric measures are more suitable for analyzing association rules. Tables 6.11 and 6.12 provide the deﬁnitions for some of these measures in terms of the frequency counts of a 2 × 2 contingency table. Consistency among Objective Measures Given the wide variety of measures available, it is reasonable to question whether the measures can produce similar ordering results when applied to a set of association patterns. If the measures are consistent, then we can choose any one of them as our evaluation metric. Otherwise, it is important to understand what their diﬀerences are in order to determine which measure is more suitable for analyzing certain types of patterns. Table 6.11. Examples of symmetric objective measures for the itemset {A, B}. Measure (Symbol)
Deﬁnition
Correlation (φ)
√N f11 −f1+ f+1 f1+ f+1 f0+ f+0 " f11 f00 f10 f01
Odds ratio (α) Kappa (κ) Interest (I) Cosine (IS)
N f11 +N f00 −f1+ f+1 −f0+ f+0 N 2 −f1+ f+1 −f0+ f+0
" N f11 f1+ f+1 " f11 f1+ f+1 −
f1+ f+1 N2
PiatetskyShapiro (P S)
f11 N
Collective strength (S)
f11 +f00 f1+ f+1 +f0+ f+0
Jaccard (ζ) Allconﬁdence (h)
×
N −f1+ f+1 −f0+ f+0 N −f11 −f00
" f1+ + f+1 − f11 11 f11 min ff1+ , f+1
f11
377
Chapter 6
Association Analysis
Table 6.12. Examples of asymmetric objective measures for the rule A −→ B. Measure (Symbol) GoodmanKruskal (λ) Mutual Information (M )
Deﬁnition " N − maxk f+k j maxk fjk − maxk f+k fij N fij " − i fNi+ log fNi+ i j N log fi+ f+j f11 N f11 f10 N f10 N log f1+ f+1 + N log f1+ f+0 f1+ f+1 2 f11 2 f10 2 N × ( f1+ ) + ( f1+ ) ] − ( N ) 0+ +0 2 01 2 00 2 + fN × [( ff0+ ) + ( ff0+ ) ] − ( fN )
JMeasure (J) Gini index (G)
Laplace (L) Conviction (V ) Certainty factor (F ) Added Value (AV )
" f11 + 1 f1+ + 2 " f1+ f+0 N f10 f11 f+1 " +1 1 − fN f1+ − N f11 f1+
−
f+1 N
Table 6.13. Example of contingency tables. Example E1 E2 E3 E4 E5 E6 E7 E8 E9 E10
f11 8123 8330 3954 2886 1500 4000 9481 4000 7450 61
f10 83 2 3080 1363 2000 2000 298 2000 2483 2483
f01 424 622 5 1320 500 1000 127 2000 4 4
f00 1370 1046 2961 4431 6000 3000 94 2000 63 7452
Suppose the symmetric and asymmetric measures are applied to rank the ten contingency tables shown in Table 6.13. These contingency tables are chosen to illustrate the diﬀerences among the existing measures. The ordering produced by these measures are shown in Tables 6.14 and 6.15, respectively (with 1 as the most interesting and 10 as the least interesting table). Although some of the measures appear to be consistent with each other, there are certain measures that produce quite diﬀerent ordering results. For example, the rankings given by the φcoeﬃcient agree with those provided by κ and collective strength, but are somewhat diﬀerent than the rankings produced by interest
378
6.7
Evaluation of Association Patterns
Table 6.14. Rankings of contingency tables using the symmetric measures given in Table 6.11. E1 E2 E3 E4 E5 E6 E7 E8 E9 E10
φ 1 2 3 4 5 6 7 8 9 10
α 3 1 2 8 7 9 6 10 4 5
κ 1 2 4 3 6 5 7 8 9 10
I 6 7 4 3 2 5 9 8 10 1
IS 2 3 5 7 9 6 1 8 4 10
PS 2 5 1 3 6 4 8 7 9 10
S 1 2 3 4 6 5 7 8 9 10
ζ 2 3 6 7 9 5 1 8 4 10
h 2 3 8 5 9 7 1 7 4 10
Table 6.15. Rankings of contingency tables using the asymmetric measures given in Table 6.12. E1 E2 E3 E4 E5 E6 E7 E8 E9 E10
λ 1 2 5 4 9 3 7 8 6 10
M 1 2 3 6 7 8 5 9 4 10
J 1 2 5 3 4 6 9 7 10 8
G 1 3 2 4 6 5 8 7 9 10
L 4 5 2 9 8 7 3 10 1 6
V 2 1 6 3 5 4 7 8 9 10
F 2 1 6 3 5 4 7 8 9 10
AV 5 6 4 1 2 3 9 7 10 8
factor and odds ratio. Furthermore, a contingency table such as E10 is ranked lowest according to the φcoeﬃcient, but highest according to interest factor. Properties of Objective Measures The results shown in Table 6.14 suggest that a signiﬁcant number of the measures provide conﬂicting information about the quality of a pattern. To understand their diﬀerences, we need to examine the properties of these measures. Inversion Property Consider the bit vectors shown in Figure 6.28. The 0/1 bit in each column vector indicates whether a transaction (row) contains a particular item (column). For example, the vector A indicates that item a
379
Chapter 6
Association Analysis A
B
C
D
E
F
1 0 0 0 0 0 0 0 0 1
0 0 0 0 1 0 0 0 0 0
0 1 1 1 1 1 1 1 1 0
1 1 1 1 0 1 1 1 1 1
0 1 1 1 1 1 1 1 1 0
0 0 0 0 1 0 0 0 0 0
(a)
(b)
(c)
Figure 6.28. Effect of the inversion operation. The vectors C and E are inversions of vector A, while the vector D is an inversion of vectors B and F .
belongs to the ﬁrst and last transactions, whereas the vector B indicates that item b is contained only in the ﬁfth transaction. The vectors C and E are in fact related to the vector A—their bits have been inverted from 0’s (absence) to 1’s (presence), and vice versa. Similarly, D is related to vectors B and F by inverting their bits. The process of ﬂipping a bit vector is called inversion. If a measure is invariant under the inversion operation, then its value for the vector pair (C, D) should be identical to its value for (A, B). The inversion property of a measure can be tested as follows. Deﬁnition 6.6 (Inversion Property). An objective measure M is invariant under the inversion operation if its value remains the same when exchanging the frequency counts f11 with f00 and f10 with f01 . Among the measures that remain invariant under this operation include the φcoeﬃcient, odds ratio, κ, and collective strength. These measures may not be suitable for analyzing asymmetric binary data. For example, the φcoeﬃcient between C and D is identical to the φcoeﬃcient between A and B, even though items c and d appear together more frequently than a and b. Furthermore, the φcoeﬃcient between C and D is less than that between E and F even though items e and f appear together only once! We had previously raised this issue when discussing the limitations of the φcoeﬃcient on page 375. For asymmetric binary data, measures that do not remain invariant under the inversion operation are preferred. Some of the noninvariant measures include interest factor, IS, P S, and the Jaccard coeﬃcient.
380
6.7
Evaluation of Association Patterns
Null Addition Property Suppose we are interested in analyzing the relationship between a pair of words, such as data and mining, in a set of documents. If a collection of articles about ice ﬁshing is added to the data set, should the association between data and mining be aﬀected? This process of adding unrelated data (in this case, documents) to a given data set is known as the null addition operation. Deﬁnition 6.7 (Null Addition Property). An objective measure M is invariant under the null addition operation if it is not aﬀected by increasing f00 , while all other frequencies in the contingency table stay the same. For applications such as document analysis or market basket analysis, the measure is expected to remain invariant under the null addition operation. Otherwise, the relationship between words may disappear simply by adding enough documents that do not contain both words! Examples of measures that satisfy this property include cosine (IS) and Jaccard (ξ) measures, while those that violate this property include interest factor, P S, odds ratio, and the φcoeﬃcient. Scaling Property Table 6.16 shows the contingency tables for gender and the grades achieved by students enrolled in a particular course in 1993 and 2004. The data in these tables showed that the number of male students has doubled since 1993, while the number of female students has increased by a factor of 3. However, the male students in 2004 are not performing any better than those in 1993 because the ratio of male students who achieve a high grade to those who achieve a low grade is still the same, i.e., 3:4. Similarly, the female students in 2004 are performing no better than those in 1993. The association between grade and gender is expected to remain unchanged despite changes in the sampling distribution. Table 6.16. The gradegender example. High Low
Male 30 40 70
Female 20 10 30
50 50 100
(a) Sample data from 1993.
High Low
Male 60 80 140
Female 60 30 90
120 110 230
(b) Sample data from 2004.
381
Chapter 6
Association Analysis Table 6.17. Properties of symmetric measures.
Symbol φ α κ I IS PS S ζ h s
Measure φcoeﬃcient odds ratio Cohen’s Interest Cosine PiatetskyShapiro’s Collective strength Jaccard Allconﬁdence Support
Inversion Yes Yes Yes No No Yes Yes No No No
Null Addition No No No No Yes No No Yes No No
Scaling No Yes No No No No No No No No
Deﬁnition 6.8 (Scaling Invariance Property). An objective measure M is invariant under the row/column scaling operation if M (T ) = M (T ), where T is a contingency table with frequency counts [f11 ; f10 ; f01 ; f00 ], T is a contingency table with scaled frequency counts [k1 k3 f11 ; k2 k3 f10 ; k1 k4 f01 ; k2 k4 f00 ], and k1 , k2 , k3 , k4 are positive constants. From Table 6.17, notice that only the odds ratio (α) is invariant under the row and column scaling operations. All other measures such as the φcoeﬃcient, κ, IS, interest factor, and collective strength (S) change their values when the rows and columns of the contingency table are rescaled. Although we do not discuss the properties of asymmetric measures (such as conﬁdence, Jmeasure, Gini index, and conviction), it is clear that such measures do not preserve their values under inversion and row/column scaling operations, but are invariant under the null addition operation.
6.7.2
Measures beyond Pairs of Binary Variables
The measures shown in Tables 6.11 and 6.12 are deﬁned for pairs of binary variables (e.g., 2itemsets or association rules). However, many of them, such as support and allconﬁdence, are also applicable to largersized itemsets. Other measures, such as interest factor, IS, P S, and Jaccard coeﬃcient, can be extended to more than two variables using the frequency tables tabulated in a multidimensional contingency table. An example of a threedimensional contingency table for a, b, and c is shown in Table 6.18. Each entry fijk in this table represents the number of transactions that contain a particular combination of items a, b, and c. For example, f101 is the number of transactions that contain a and c, but not b. On the other hand, a marginal frequency
382
6.7
Evaluation of Association Patterns
Table 6.18. Example of a threedimensional contingency table.
c
b
b
c
b
b
a
f111
f101
f1+1
a
f110
f100
f1+0
a
f011
f001
f0+1
a
f010
f000
f0+0
f+11
f+01
f++1
f+10
f+00
f++0
such as f1+1 is the number of transactions that contain a and c, irrespective of whether b is present in the transaction. Given a kitemset {i1 , i2 , . . . , ik }, the condition for statistical independence can be stated as follows: fi1 i2 ...ik =
fi1 +...+ × f+i2 ...+ × . . . × f++...ik . N k−1
(6.12)
With this deﬁnition, we can extend objective measures such as interest factor and P S, which are based on deviations from statistical independence, to more than two variables: I = PS =
N k−1 × fi1 i2 ...ik fi1 +...+ × f+i2 ...+ × . . . × f++...ik fi +...+ × f+i2 ...+ × . . . × f++...ik fi1 i2 ...ik − 1 N Nk
Another approach is to deﬁne the objective measure as the maximum, minimum, or average value for the associations between pairs of items in a pattern. For example, given a kitemset X = {i1 , i2 , . . . , ik }, we may deﬁne the φcoeﬃcient for X as the average φcoeﬃcient between every pair of items (ip , iq ) in X. However, because the measure considers only pairwise associations, it may not capture all the underlying relationships within a pattern. Analysis of multidimensional contingency tables is more complicated because of the presence of partial associations in the data. For example, some associations may appear or disappear when conditioned upon the value of certain variables. This problem is known as Simpson’s paradox and is described in the next section. More sophisticated statistical techniques are available to analyze such relationships, e.g., loglinear models, but these techniques are beyond the scope of this book.
383
Chapter 6
Association Analysis
Table 6.19. A twoway contingency table between the sale of highdefinition television and exercise machine. Buy Buy Exercise Machine HDTV Yes No Yes 99 81 180 No 54 66 120 153 147 300
Table 6.20. Example of a threeway contingency table. Customer Group College Students Working Adult
6.7.3
Buy HDTV Yes No Yes No
Buy Exercise Yes 1 4 98 50
Machine No 9 30 72 36
Total 10 34 170 86
Simpson’s Paradox
It is important to exercise caution when interpreting the association between variables because the observed relationship may be inﬂuenced by the presence of other confounding factors, i.e., hidden variables that are not included in the analysis. In some cases, the hidden variables may cause the observed relationship between a pair of variables to disappear or reverse its direction, a phenomenon that is known as Simpson’s paradox. We illustrate the nature of this paradox with the following example. Consider the relationship between the sale of highdeﬁnition television (HDTV) and exercise machine, as shown in Table 6.19. The rule {HDTV=Yes} −→ {Exercise machine=Yes} has a conﬁdence of 99/180 = 55% and the rule {HDTV=No} −→ {Exercise machine=Yes} has a conﬁdence of 54/120 = 45%. Together, these rules suggest that customers who buy highdeﬁnition televisions are more likely to buy exercise machines than those who do not buy highdeﬁnition televisions. However, a deeper analysis reveals that the sales of these items depend on whether the customer is a college student or a working adult. Table 6.20 summarizes the relationship between the sale of HDTVs and exercise machines among college students and working adults. Notice that the support counts given in the table for college students and working adults sum up to the frequencies shown in Table 6.19. Furthermore, there are more working adults
384
6.7
Evaluation of Association Patterns
than college students who buy these items. For college students: c {HDTV=Yes} −→ {Exercise machine=Yes} = 1/10 = 10%, c {HDTV=No} −→ {Exercise machine=Yes} = 4/34 = 11.8%, while for working adults: c {HDTV=Yes} −→ {Exercise machine=Yes} = 98/170 = 57.7%, c {HDTV=No} −→ {Exercise machine=Yes} = 50/86 = 58.1%. The rules suggest that, for each group, customers who do not buy highdeﬁnition televisions are more likely to buy exercise machines, which contradict the previous conclusion when data from the two customer groups are pooled together. Even if alternative measures such as correlation, odds ratio, or interest are applied, we still ﬁnd that the sale of HDTV and exercise machine is positively correlated in the combined data but is negatively correlated in the stratiﬁed data (see Exercise 20 on page 414). The reversal in the direction of association is known as Simpson’s paradox. The paradox can be explained in the following way. Notice that most customers who buy HDTVs are working adults. Working adults are also the largest group of customers who buy exercise machines. Because nearly 85% of the customers are working adults, the observed relationship between HDTV and exercise machine turns out to be stronger in the combined data than what it would have been if the data is stratiﬁed. This can also be illustrated mathematically as follows. Suppose a/b < c/d and p/q < r/s, where a/b and p/q may represent the conﬁdence of the rule A −→ B in two diﬀerent strata, while c/d and r/s may represent the conﬁdence of the rule A −→ B in the two strata. When the data is pooled together, the conﬁdence values of the rules in the combined data are (a + p)/(b + q) and (c + r)/(d + s), respectively. Simpson’s paradox occurs when c+r a+p > , b+q d+s thus leading to the wrong conclusion about the relationship between the variables. The lesson here is that proper stratiﬁcation is needed to avoid generating spurious patterns resulting from Simpson’s paradox. For example, market
385
Chapter 6
Association Analysis ×10 4
5 4.5 4 3.5
Support
3 2.5 2 1.5 1 0.5 0
0
500
1000 1500 Items sorted by support
2000
2500
Figure 6.29. Support distribution of items in the census data set.
basket data from a major supermarket chain should be stratiﬁed according to store locations, while medical records from various patients should be stratiﬁed according to confounding factors such as age and gender.
6.8
Eﬀect of Skewed Support Distribution
The performances of many association analysis algorithms are inﬂuenced by properties of their input data. For example, the computational complexity of the Apriori algorithm depends on properties such as the number of items in the data and average transaction width. This section examines another important property that has signiﬁcant inﬂuence on the performance of association analysis algorithms as well as the quality of extracted patterns. More speciﬁcally, we focus on data sets with skewed support distributions, where most of the items have relatively low to moderate frequencies, but a small number of them have very high frequencies. An example of a real data set that exhibits such a distribution is shown in Figure 6.29. The data, taken from the PUMS (Public Use Microdata Sample) census data, contains 49,046 records and 2113 asymmetric binary variables. We shall treat the asymmetric binary variables as items and records as transactions in the remainder of this section. While more than 80% of the items have support less than 1%, a handful of them have support greater than 90%.
386
6.8
Eﬀect of Skewed Support Distribution
Table 6.21. Grouping the items in the census data set based on their support values. Group Support Number of Items
G1 < 1% 1735
G2 1% − 90% 358
G3 > 90% 20
To illustrate the eﬀect of skewed support distribution on frequent itemset mining, we divide the items into three groups, G1 , G2 , and G3 , according to their support levels. The number of items that belong to each group is shown in Table 6.21. Choosing the right support threshold for mining this data set can be quite tricky. If we set the threshold too high (e.g., 20%), then we may miss many interesting patterns involving the low support items from G1 . In market basket analysis, such low support items may correspond to expensive products (such as jewelry) that are seldom bought by customers, but whose patterns are still interesting to retailers. Conversely, when the threshold is set too low, it becomes diﬃcult to ﬁnd the association patterns due to the following reasons. First, the computational and memory requirements of existing association analysis algorithms increase considerably with low support thresholds. Second, the number of extracted patterns also increases substantially with low support thresholds. Third, we may extract many spurious patterns that relate a highfrequency item such as milk to a lowfrequency item such as caviar. Such patterns, which are called crosssupport patterns, are likely to be spurious because their correlations tend to be weak. For example, at a support threshold equal to 0.05%, there are 18,847 frequent pairs involving items from G1 and G3 . Out of these, 93% of them are crosssupport patterns; i.e., the patterns contain items from both G1 and G3 . The maximum correlation obtained from the crosssupport patterns is 0.029, which is much lower than the maximum correlation obtained from frequent patterns involving items from the same group (which is as high as 1.0). Similar statement can be made about many other interestingness measures discussed in the previous section. This example shows that a large number of weakly correlated crosssupport patterns can be generated when the support threshold is suﬃciently low. Before presenting a methodology for eliminating such patterns, we formally deﬁne the concept of crosssupport patterns.
387
Chapter 6
Association Analysis
Deﬁnition 6.9 (CrossSupport Pattern). A crosssupport pattern is an itemset X = {i1 , i2 , . . . , ik } whose support ratio min s(i1 ), s(i2 ), . . . , s(ik ) , r(X) = (6.13) max s(i1 ), s(i2 ), . . . , s(ik ) is less than a userspeciﬁed threshold hc . Example 6.4. Suppose the support for milk is 70%, while the support for sugar is 10% and caviar is 0.04%. Given hc = 0.01, the frequent itemset {milk, sugar, caviar} is a crosssupport pattern because its support ratio is min 0.7, 0.1, 0.0004 0.0004 = = 0.00058 < 0.01. r= 0.7 max 0.7, 0.1, 0.0004
Existing measures such as support and conﬁdence may not be suﬃcient to eliminate crosssupport patterns, as illustrated by the data set shown in Figure 6.30. Assuming that hc = 0.3, the itemsets {p, q}, {p, r}, and {p, q, r} are crosssupport patterns because their support ratios, which are equal to 0.2, are less than the threshold hc . Although we can apply a high support threshold, say, 20%, to eliminate the crosssupport patterns, this may come at the expense of discarding other interesting patterns such as the strongly correlated itemset, {q, r} that has support equal to 16.7%. Conﬁdence pruning also does not help because the conﬁdence of the rules extracted from crosssupport patterns can be very high. For example, the conﬁdence for {q} −→ {p} is 80% even though {p, q} is a crosssupport pattern. The fact that the crosssupport pattern can produce a highconﬁdence rule should not come as a surprise because one of its items (p) appears very frequently in the data. Therefore, p is expected to appear in many of the transactions that contain q. Meanwhile, the rule {q} −→ {r} also has high conﬁdence even though {q, r} is not a crosssupport pattern. This example demonstrates the diﬃculty of using the conﬁdence measure to distinguish between rules extracted from crosssupport and noncrosssupport patterns. Returning to the previous example, notice that the rule {p} −→ {q} has very low conﬁdence because most of the transactions that contain p do not contain q. In contrast, the rule {r} −→ {q}, which is derived from the pattern {q, r}, has very high conﬁdence. This observation suggests that crosssupport patterns can be detected by examining the lowest conﬁdence rule that can be extracted from a given itemset. The proof of this statement can be understood as follows.
388
6.8
Eﬀect of Skewed Support Distribution p 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 0 0 0 0
q 1 1 1 1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
r 1 1 1 1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
Figure 6.30. A transaction data set containing three items, p, q, and r, where p is a high support item and q and r are low support items.
1. Recall the following antimonotone property of conﬁdence: conf ({i1 i2 } −→ {i3 , i4 , . . . , ik }) ≤ conf ({i1 i2 i3 } −→ {i4 , i5 , . . . , ik }). This property suggests that conﬁdence never increases as we shift more items from the left to the righthand side of an association rule. Because of this property, the lowest conﬁdence rule extracted from a frequent itemset contains only one item on its lefthand side. We denote the set of all rules with only one item on its lefthand side as R1 . 2. Given a frequent itemset {i1 , i2 , . . . , ik }, the rule {ij } −→ {i1 , i2 , . . . , ij−1 , ij+1 , . . . , ik } has the lowest conﬁdence in R1 if s(ij ) = max s(i1 ), s(i2 ), . . . , s(ik ) . This follows directly from the deﬁnition of conﬁdence as the ratio between the rule’s support and the support of the rule antecedent.
389
Chapter 6
Association Analysis
3. Summarizing the previous points, the lowest conﬁdence attainable from a frequent itemset {i1 , i2 , . . . , ik } is s({i , i , . . . , ik }) 1 2 . max s(i1 ), s(i2 ), . . . , s(ik ) This expression is also known as the hconﬁdence or allconﬁdence measure. Because of the antimonotone property of support, the numerator of the hconﬁdence measure is bounded by the minimum support of any item that appears in the frequent itemset. In other words, the hconﬁdence of an itemset X = {i1 , i2 , . . . , ik } must not exceed the following expression: min s(i1 ), s(i2 ), . . . , s(ik ) . hconﬁdence(X) ≤ max s(i1 ), s(i2 ), . . . , s(ik ) Note the equivalence between the upper bound of hconﬁdence and the support ratio (r) given in Equation 6.13. Because the support ratio for a crosssupport pattern is always less than hc , the hconﬁdence of the pattern is also guaranteed to be less than hc . Therefore, crosssupport patterns can be eliminated by ensuring that the hconﬁdence values for the patterns exceed hc . As a ﬁnal note, it is worth mentioning that the advantages of using hconﬁdence go beyond eliminating crosssupport patterns. The measure is also antimonotone, i.e., hconﬁdence({i1 , i2 , . . . , ik }) ≥ hconﬁdence({i1 , i2 , . . . , ik+1 }), and thus can be incorporated directly into the mining algorithm. Furthermore, hconﬁdence ensures that the items contained in an itemset are strongly associated with each other. For example, suppose the hconﬁdence of an itemset X is 80%. If one of the items in X is present in a transaction, there is at least an 80% chance that the rest of the items in X also belong to the same transaction. Such strongly associated patterns are called hyperclique patterns.
6.9
Bibliographic Notes
The association rule mining task was ﬁrst introduced by Agrawal et al. in [228, 229] to discover interesting relationships among items in market basket
390
6.9
Bibliographic Notes
transactions. Since its inception, extensive studies have been conducted to address the various conceptual, implementation, and application issues pertaining to the association analysis task. A summary of the various research activities in this area is shown in Figure 6.31. Conceptual Issues Research in conceptual issues is focused primarily on (1) developing a framework to describe the theoretical underpinnings of association analysis, (2) extending the formulation to handle new types of patterns, and (3) extending the formulation to incorporate attribute types beyond asymmetric binary data. Following the pioneering work by Agrawal et al., there has been a vast amount of research on developing a theory for the association analysis problem. In [254], Gunopoulos et al. showed a relation between the problem of ﬁnding maximal frequent itemsets and the hypergraph transversal problem. An upper bound on the complexity of association analysis task was also derived. Zaki et al. [334, 336] and Pasquier et al. [294] have applied formal concept analysis to study the frequent itemset generation problem. The work by Zaki et al. have subsequently led them to introduce the notion of closed frequent itemsets [336]. Friedman et al. have studied the association analysis problem in the context of bump hunting in multidimensional space [252]. More speciﬁcally, they consider frequent itemset generation as the task of ﬁnding high probability density regions in multidimensional space. Over the years, new types of patterns have been deﬁned, such as proﬁle association rules [225], cyclic association rules [290], fuzzy association rules [273], exception rules [316], negative association rules [238, 304], weighted association rules [240, 300], dependence rules [308], peculiar rules[340], intertransaction association rules [250, 323], and partial classiﬁcation rules [231, 285]. Other types of patterns include closed itemsets [294, 336], maximal itemsets [234], hyperclique patterns [330], support envelopes [314], emerging patterns [246], and contrast sets [233]. Association analysis has also been successfully applied to sequential [230, 312], spatial [266], and graphbased [268, 274, 293, 331, 335] data. The concept of crosssupport pattern was ﬁrst introduced by Hui et al. in [330]. An eﬃcient algorithm (called Hyperclique Miner) that automatically eliminates crosssupport patterns was also proposed by the authors. Substantial research has been conducted to extend the original association rule formulation to nominal [311], ordinal [281], interval [284], and ratio [253, 255, 311, 325, 339] attributes. One of the key issues is how to deﬁne the support measure for these attributes. A methodology was proposed by Steinbach et
391
Chapter 6
Association Analysis
al. [315] to extend the traditional notion of support to more general patterns and attribute types.
392
negative dependence causal weighted spatial and colocation patterns temporal (cyclic, sequential) fuzzy exception rules
Rules
lattice theory bounds on itemset enumeration
Theroretical Formulation
subtrees subgraphs
Other Structures
binary numeric nominal ordinal mixed
Data Type
serial or parallel online or batch
Apriori DIC treeprojeciton FPtree Hmine Partition Samplingbased CHARM
Postprocessing
objective subjective
Web analysis text analysis bioinformatics Earth Science
Domains
ranking filtering summarizing
Method
classification regression clustering recommender systems
Other data mining problems
Application Issues
Interestingness
Measure
Visualization
item taxonomy templatebased multiple support
Constraints
Algorithm and Data Structure
Pattern Discovery
Computational model
optimization SQL support OLAP multidatabase
Database issues
Implementation Issues
Figure 6.31. A summary of the various research activities in association analysis.
closed maximal emerging patterns hyperclique patterns support envelope
Itemsets
Type of Patterns
Conceptual Issues
Research Issues in Mining Association Patterns
6.9 Bibliographic Notes
393
Chapter 6
Association Analysis
Implementation Issues Research activities in this area revolve around (1) integrating the mining capability into existing database technology, (2) developing eﬃcient and scalable mining algorithms, (3) handling userspeciﬁed or domainspeciﬁc constraints, and (4) postprocessing the extracted patterns. There are several advantages to integrating association analysis into existing database technology. First, it can make use of the indexing and query processing capabilities of the database system. Second, it can also exploit the DBMS support for scalability, checkpointing, and parallelization [301]. The SETM algorithm developed by Houtsma et al. [265] was one of the earliest algorithms to support association rule discovery via SQL queries. Since then, numerous methods have been developed to provide capabilities for mining association rules in database systems. For example, the DMQL [258] and MSQL [267] query languages extend the basic SQL with new operators for mining association rules. The Mine Rule operator [283] is an expressive SQL operator that can handle both clustered attributes and item hierarchies. Tsur et al. [322] developed a generateandtest approach called query ﬂocks for mining association rules. A distributed OLAPbased infrastructure was developed by Chen et al. [241] for mining multilevel association rules. Dunkel and Soparkar [248] investigated the time and storage complexity of the Apriori algorithm. The FPgrowth algorithm was developed by Han et al. in [259]. Other algorithms for mining frequent itemsets include the DHP (dynamic hashing and pruning) algorithm proposed by Park et al. [292] and the Partition algorithm developed by Savasere et al [303]. A samplingbased frequent itemset generation algorithm was proposed by Toivonen [320]. The algorithm requires only a single pass over the data, but it can produce more candidate itemsets than necessary. The Dynamic Itemset Counting (DIC) algorithm [239] makes only 1.5 passes over the data and generates less candidate itemsets than the samplingbased algorithm. Other notable algorithms include the treeprojection algorithm [223] and HMine [295]. Survey articles on frequent itemset generation algorithms can be found in [226, 262]. A repository of data sets and algorithms is available at the Frequent Itemset Mining Implementations (FIMI) repository (http://ﬁmi.cs.helsinki.ﬁ). Parallel algorithms for mining association patterns have been developed by various authors [224, 256, 287, 306, 337]. A survey of such algorithms can be found in [333]. Online and incremental versions of association rule mining algorithms had also been proposed by Hidber [260] and Cheung et al. [242]. Srikant et al. [313] have considered the problem of mining association rules in the presence of boolean constraints such as the following:
394
6.9
Bibliographic Notes
(Cookies ∧ Milk) ∨ (descendents(Cookies) ∧ ¬ancestors(Wheat Bread)) Given such a constraint, the algorithm looks for rules that contain both cookies and milk, or rules that contain the descendent items of cookies but not ancestor items of wheat bread. Singh et al. [310] and Ng et al. [288] had also developed alternative techniques for constrainedbased association rule mining. Constraints can also be imposed on the support for diﬀerent itemsets. This problem was investigated by Wang et al. [324], Liu et al. in [279], and Seno et al. [305]. One potential problem with association analysis is the large number of patterns that can be generated by current algorithms. To overcome this problem, methods to rank, summarize, and ﬁlter patterns have been developed. Toivonen et al. [321] proposed the idea of eliminating redundant rules using structural rule covers and to group the remaining rules using clustering. Liu et al. [280] applied the statistical chisquare test to prune spurious patterns and summarized the remaining patterns using a subset of the patterns called direction setting rules. The use of objective measures to ﬁlter patterns has been investigated by many authors, including Brin et al. [238], Bayardo and Agrawal [235], Aggarwal and Yu [227], and DuMouchel and Pregibon[247]. The properties for many of these measures were analyzed by PiatetskyShapiro [297], Kamber and Singhal [270], Hilderman and Hamilton [261], and Tan et al. [318]. The gradegender example used to highlight the importance of the row and column scaling invariance property was heavily inﬂuenced by the discussion given in [286] by Mosteller. Meanwhile, the teacoﬀee example illustrating the limitation of conﬁdence was motivated by an example given in [238] by Brin et al. Because of the limitation of conﬁdence, Brin et al. [238] had proposed the idea of using interest factor as a measure of interestingness. The allconﬁdence measure was proposed by Omiecinski [289]. Xiong et al. [330] introduced the crosssupport property and showed that the allconﬁdence measure can be used to eliminate crosssupport patterns. A key diﬃculty in using alternative objective measures besides support is their lack of a monotonicity property, which makes it diﬃcult to incorporate the measures directly into the mining algorithms. Xiong et al. [328] have proposed an eﬃcient method for mining correlations by introducing an upper bound function to the φcoeﬃcient. Although the measure is nonmonotone, it has an upper bound expression that can be exploited for the eﬃcient mining of strongly correlated itempairs. Fabris and Freitas [249] have proposed a method for discovering interesting associations by detecting the occurrences of Simpson’s paradox [309]. Megiddo and Srikant [282] described an approach for validating the extracted
395
Chapter 6
Association Analysis
patterns using hypothesis testing methods. A resamplingbased technique was also developed to avoid generating spurious patterns because of the multiple comparison problem. Bolton et al. [237] have applied the BenjaminiHochberg [236] and Bonferroni correction methods to adjust the pvalues of discovered patterns in market basket data. Alternative methods for handling the multiple comparison problem were suggested by Webb [326] and Zhang et al. [338]. Application of subjective measures to association analysis has been investigated by many authors. Silberschatz and Tuzhilin [307] presented two principles in which a rule can be considered interesting from a subjective point of view. The concept of unexpected condition rules was introduced by Liu et al. in [277]. Cooley et al. [243] analyzed the idea of combining soft belief sets using the DempsterShafer theory and applied this approach to identify contradictory and novel association patterns in Web data. Alternative approaches include using Bayesian networks [269] and neighborhoodbased information [245] to identify subjectively interesting patterns. Visualization also helps the user to quickly grasp the underlying structure of the discovered patterns. Many commercial data mining tools display the complete set of rules (which satisfy both support and conﬁdence threshold criteria) as a twodimensional plot, with each axis corresponding to the antecedent or consequent itemsets of the rule. Hofmann et al. [263] proposed using Mosaic plots and Double Decker plots to visualize association rules. This approach can visualize not only a particular rule, but also the overall contingency table between itemsets in the antecedent and consequent parts of the rule. Nevertheless, this technique assumes that the rule consequent consists of only a single attribute. Application Issues Association analysis has been applied to a variety of application domains such as Web mining [296, 317], document analysis [264], telecommunication alarm diagnosis [271], network intrusion detection [232, 244, 275], and bioinformatics [302, 327]. Applications of association and correlation pattern analysis to Earth Science studies have been investigated in [298, 299, 319]. Association patterns have also been applied to other learning problems such as classiﬁcation [276, 278], regression [291], and clustering [257, 329, 332]. A comparison between classiﬁcation and association rule mining was made by Freitas in his position paper [251]. The use of association patterns for clustering has been studied by many authors including Han et al.[257], Kosters et al. [272], Yang et al. [332] and Xiong et al. [329].
396
Bibliography
Bibliography [223] R. C. Agarwal, C. C. Aggarwal, and V. V. V. Prasad. A Tree Projection Algorithm for Generation of Frequent Itemsets. Journal of Parallel and Distributed Computing (Special Issue on High Performance Data Mining), 61(3):350–371, 2001. [224] R. C. Agarwal and J. C. Shafer. Parallel Mining of Association Rules. IEEE Transactions on Knowledge and Data Engineering, 8(6):962–969, March 1998. [225] C. C. Aggarwal, Z. Sun, and P. S. Yu. Online Generation of Proﬁle Association Rules. In Proc. of the 4th Intl. Conf. on Knowledge Discovery and Data Mining, pages 129– 133, New York, NY, August 1996. [226] C. C. Aggarwal and P. S. Yu. Mining Large Itemsets for Association Rules. Data Engineering Bulletin, 21(1):23–31, March 1998. [227] C. C. Aggarwal and P. S. Yu. Mining Associations with the Collective Strength Approach. IEEE Trans. on Knowledge and Data Engineering, 13(6):863–873, January/February 2001. [228] R. Agrawal, T. Imielinski, and A. Swami. Database mining: A performance perspective. IEEE Transactions on Knowledge and Data Engineering, 5:914–925, 1993. [229] R. Agrawal, T. Imielinski, and A. Swami. Mining association rules between sets of items in large databases. In Proc. ACM SIGMOD Intl. Conf. Management of Data, pages 207–216, Washington, DC, 1993. [230] R. Agrawal and R. Srikant. Mining Sequential Patterns. In Proc. of Intl. Conf. on Data Engineering, pages 3–14, Taipei, Taiwan, 1995. [231] K. Ali, S. Manganaris, and R. Srikant. Partial Classiﬁcation using Association Rules. In Proc. of the 3rd Intl. Conf. on Knowledge Discovery and Data Mining, pages 115– 118, Newport Beach, CA, August 1997. [232] D. Barbar´ a, J. Couto, S. Jajodia, and N. Wu. ADAM: A Testbed for Exploring the Use of Data Mining in Intrusion Detection. SIGMOD Record, 30(4):15–24, 2001. [233] S. D. Bay and M. Pazzani. Detecting Group Diﬀerences: Mining Contrast Sets. Data Mining and Knowledge Discovery, 5(3):213–246, 2001. [234] R. Bayardo. Eﬃciently Mining Long Patterns from Databases. In Proc. of 1998 ACMSIGMOD Intl. Conf. on Management of Data, pages 85–93, Seattle, WA, June 1998. [235] R. Bayardo and R. Agrawal. Mining the Most Interesting Rules. In Proc. of the 5th Intl. Conf. on Knowledge Discovery and Data Mining, pages 145–153, San Diego, CA, August 1999. [236] Y. Benjamini and Y. Hochberg. Controlling the False Discovery Rate: A Practical and Powerful Approach to Multiple Testing. Journal Royal Statistical Society B, 57 (1):289–300, 1995. [237] R. J. Bolton, D. J. Hand, and N. M. Adams. Determining Hit Rate in Pattern Search. In Proc. of the ESF Exploratory Workshop on Pattern Detection and Discovery in Data Mining, pages 36–48, London, UK, September 2002. [238] S. Brin, R. Motwani, and C. Silverstein. Beyond market baskets: Generalizing association rules to correlations. In Proc. ACM SIGMOD Intl. Conf. Management of Data, pages 265–276, Tucson, AZ, 1997. [239] S. Brin, R. Motwani, J. Ullman, and S. Tsur. Dynamic Itemset Counting and Implication Rules for market basket data. In Proc. of 1997 ACMSIGMOD Intl. Conf. on Management of Data, pages 255–264, Tucson, AZ, June 1997. [240] C. H. Cai, A. Fu, C. H. Cheng, and W. W. Kwong. Mining Association Rules with Weighted Items. In Proc. of IEEE Intl. Database Engineering and Applications Symp., pages 68–77, Cardiﬀ, Wales, 1998.
397
Chapter 6
Association Analysis
[241] Q. Chen, U. Dayal, and M. Hsu. A Distributed OLAP infrastructure for ECommerce. In Proc. of the 4th IFCIS Intl. Conf. on Cooperative Information Systems, pages 209– 220, Edinburgh, Scotland, 1999. [242] D. C. Cheung, S. D. Lee, and B. Kao. A General Incremental Technique for Maintaining Discovered Association Rules. In Proc. of the 5th Intl. Conf. on Database Systems for Advanced Applications, pages 185–194, Melbourne, Australia, 1997. [243] R. Cooley, P. N. Tan, and J. Srivastava. Discovery of Interesting Usage Patterns from Web Data. In M. Spiliopoulou and B. Masand, editors, Advances in Web Usage Analysis and User Profiling, volume 1836, pages 163–182. Lecture Notes in Computer Science, 2000. [244] P. Dokas, L. Ert¨ oz, V. Kumar, A. Lazarevic, J. Srivastava, and P. N. Tan. Data Mining for Network Intrusion Detection. In Proc. NSF Workshop on Next Generation Data Mining, Baltimore, MD, 2002. [245] G. Dong and J. Li. Interestingness of discovered association rules in terms of neighborhoodbased unexpectedness. In Proc. of the 2nd PacificAsia Conf. on Knowledge Discovery and Data Mining, pages 72–86, Melbourne, Australia, April 1998. [246] G. Dong and J. Li. Eﬃcient Mining of Emerging Patterns: Discovering Trends and Diﬀerences. In Proc. of the 5th Intl. Conf. on Knowledge Discovery and Data Mining, pages 43–52, San Diego, CA, August 1999. [247] W. DuMouchel and D. Pregibon. Empirical Bayes Screening for MultiItem Associations. In Proc. of the 7th Intl. Conf. on Knowledge Discovery and Data Mining, pages 67–76, San Francisco, CA, August 2001. [248] B. Dunkel and N. Soparkar. Data Organization and Access for Eﬃcient Data Mining. In Proc. of the 15th Intl. Conf. on Data Engineering, pages 522–529, Sydney, Australia, March 1999. [249] C. C. Fabris and A. A. Freitas. Discovering surprising patterns by detecting occurrences of Simpson’s paradox. In Proc. of the 19th SGES Intl. Conf. on KnowledgeBased Systems and Applied Artificial Intelligence), pages 148–160, Cambridge, UK, December 1999. [250] L. Feng, H. J. Lu, J. X. Yu, and J. Han. Mining intertransaction associations with templates. In Proc. of the 8th Intl. Conf. on Information and Knowledge Management, pages 225–233, Kansas City, Missouri, Nov 1999. [251] A. A. Freitas. Understanding the crucial diﬀerences between classiﬁcation and discovery of association rules—a position paper. SIGKDD Explorations, 2(1):65–69, 2000. [252] J. H. Friedman and N. I. Fisher. Bump hunting in highdimensional data. Statistics and Computing, 9(2):123–143, April 1999. [253] T. Fukuda, Y. Morimoto, S. Morishita, and T. Tokuyama. Mining Optimized Association Rules for Numeric Attributes. In Proc. of the 15th Symp. on Principles of Database Systems, pages 182–191, Montreal, Canada, June 1996. [254] D. Gunopulos, R. Khardon, H. Mannila, and H. Toivonen. Data Mining, Hypergraph Transversals, and Machine Learning. In Proc. of the 16th Symp. on Principles of Database Systems, pages 209–216, Tucson, AZ, May 1997. [255] E.H. Han, G. Karypis, and V. Kumar. MinApriori: An Algorithm for Finding Association Rules in Data with Continuous Attributes. http://www.cs.umn.edu/˜han, 1997. [256] E.H. Han, G. Karypis, and V. Kumar. Scalable Parallel Data Mining for Association Rules. In Proc. of 1997 ACMSIGMOD Intl. Conf. on Management of Data, pages 277–288, Tucson, AZ, May 1997.
398
Bibliography [257] E.H. Han, G. Karypis, V. Kumar, and B. Mobasher. Clustering Based on Association Rule Hypergraphs. In Proc. of the 1997 ACM SIGMOD Workshop on Research Issues in Data Mining and Knowledge Discovery, Tucson, AZ, 1997. [258] J. Han, Y. Fu, K. Koperski, W. Wang, and O. R. Za¨ıane. DMQL: A data mining query language for relational databases. In Proc. of the 1996 ACM SIGMOD Workshop on Research Issues in Data Mining and Knowledge Discovery, Montreal, Canada, June 1996. [259] J. Han, J. Pei, and Y. Yin. Mining Frequent Patterns without Candidate Generation. In Proc. ACMSIGMOD Int. Conf. on Management of Data (SIGMOD’00), pages 1–12, Dallas, TX, May 2000. [260] C. Hidber. Online Association Rule Mining. In Proc. of 1999 ACMSIGMOD Intl. Conf. on Management of Data, pages 145–156, Philadelphia, PA, 1999. [261] R. J. Hilderman and H. J. Hamilton. Knowledge Discovery and Measures of Interest. Kluwer Academic Publishers, 2001. [262] J. Hipp, U. Guntzer, and G. Nakhaeizadeh. Algorithms for Association Rule Mining— A General Survey. SigKDD Explorations, 2(1):58–64, June 2000. [263] H. Hofmann, A. P. J. M. Siebes, and A. F. X. Wilhelm. Visualizing Association Rules with Interactive Mosaic Plots. In Proc. of the 6th Intl. Conf. on Knowledge Discovery and Data Mining, pages 227–235, Boston, MA, August 2000. [264] J. D. Holt and S. M. Chung. Eﬃcient Mining of Association Rules in Text Databases. In Proc. of the 8th Intl. Conf. on Information and Knowledge Management, pages 234–242, Kansas City, Missouri, 1999. [265] M. Houtsma and A. Swami. Setoriented Mining for Association Rules in Relational Databases. In Proc. of the 11th Intl. Conf. on Data Engineering, pages 25–33, Taipei, Taiwan, 1995. [266] Y. Huang, S. Shekhar, and H. Xiong. Discovering Colocation Patterns from Spatial Datasets: A General Approach. IEEE Trans. on Knowledge and Data Engineering, 16 (12):1472–1485, December 2004. [267] T. Imielinski, A. Virmani, and A. Abdulghani. DataMine: Application Programming Interface and Query Language for Database Mining. In Proc. of the 2nd Intl. Conf. on Knowledge Discovery and Data Mining, pages 256–262, Portland, Oregon, 1996. [268] A. Inokuchi, T. Washio, and H. Motoda. An Aprioribased Algorithm for Mining Frequent Substructures from Graph Data. In Proc. of the 4th European Conf. of Principles and Practice of Knowledge Discovery in Databases, pages 13–23, Lyon, France, 2000. [269] S. Jaroszewicz and D. Simovici. Interestingness of Frequent Itemsets Using Bayesian Networks as Background Knowledge. In Proc. of the 10th Intl. Conf. on Knowledge Discovery and Data Mining, pages 178–186, Seattle, WA, August 2004. [270] M. Kamber and R. Shinghal. Evaluating the Interestingness of Characteristic Rules. In Proc. of the 2nd Intl. Conf. on Knowledge Discovery and Data Mining, pages 263–266, Portland, Oregon, 1996. [271] M. Klemettinen. A Knowledge Discovery Methodology for Telecommunication Network Alarm Databases. PhD thesis, University of Helsinki, 1999. [272] W. A. Kosters, E. Marchiori, and A. Oerlemans. Mining Clusters with Association Rules. In The 3rd Symp. on Intelligent Data Analysis (IDA99), pages 39–50, Amsterdam, August 1999. [273] C. M. Kuok, A. Fu, and M. H. Wong. Mining Fuzzy Association Rules in Databases. ACM SIGMOD Record, 27(1):41–46, March 1998.
399
Chapter 6
Association Analysis
[274] M. Kuramochi and G. Karypis. Frequent Subgraph Discovery. In Proc. of the 2001 IEEE Intl. Conf. on Data Mining, pages 313–320, San Jose, CA, November 2001. [275] W. Lee, S. J. Stolfo, and K. W. Mok. Adaptive Intrusion Detection: A Data Mining Approach. Artificial Intelligence Review, 14(6):533–567, 2000. [276] W. Li, J. Han, and J. Pei. CMAR: Accurate and Eﬃcient Classiﬁcation Based on Multiple Classassociation Rules. In Proc. of the 2001 IEEE Intl. Conf. on Data Mining, pages 369–376, San Jose, CA, 2001. [277] B. Liu, W. Hsu, and S. Chen. Using General Impressions to Analyze Discovered Classiﬁcation Rules. In Proc. of the 3rd Intl. Conf. on Knowledge Discovery and Data Mining, pages 31–36, Newport Beach, CA, August 1997. [278] B. Liu, W. Hsu, and Y. Ma. Integrating Classiﬁcation and Association Rule Mining. In Proc. of the 4th Intl. Conf. on Knowledge Discovery and Data Mining, pages 80–86, New York, NY, August 1998. [279] B. Liu, W. Hsu, and Y. Ma. Mining association rules with multiple minimum supports. In Proc. of the 5th Intl. Conf. on Knowledge Discovery and Data Mining, pages 125– 134, San Diego, CA, August 1999. [280] B. Liu, W. Hsu, and Y. Ma. Pruning and Summarizing the Discovered Associations. In Proc. of the 5th Intl. Conf. on Knowledge Discovery and Data Mining, pages 125–134, San Diego, CA, August 1999. [281] A. Marcus, J. I. Maletic, and K.I. Lin. Ordinal association rules for error identiﬁcation in data sets. In Proc. of the 10th Intl. Conf. on Information and Knowledge Management, pages 589–591, Atlanta, GA, October 2001. [282] N. Megiddo and R. Srikant. Discovering Predictive Association Rules. In Proc. of the 4th Intl. Conf. on Knowledge Discovery and Data Mining, pages 274–278, New York, August 1998. [283] R. Meo, G. Psaila, and S. Ceri. A New SQLlike Operator for Mining Association Rules. In Proc. of the 22nd VLDB Conf., pages 122–133, Bombay, India, 1996. [284] R. J. Miller and Y. Yang. Association Rules over Interval Data. In Proc. of 1997 ACMSIGMOD Intl. Conf. on Management of Data, pages 452–461, Tucson, AZ, May 1997. [285] Y. Morimoto, T. Fukuda, H. Matsuzawa, T. Tokuyama, and K. Yoda. Algorithms for mining association rules for binary segmentations of huge categorical databases. In Proc. of the 24th VLDB Conf., pages 380–391, New York, August 1998. [286] F. Mosteller. Association and Estimation in Contingency Tables. Journal of the American Statistical Association, 63:1–28, 1968. [287] A. Mueller. Fast sequential and parallel algorithms for association rule mining: A comparison. Technical Report CSTR3515, University of Maryland, August 1995. [288] R. T. Ng, L. V. S. Lakshmanan, J. Han, and A. Pang. Exploratory Mining and Pruning Optimizations of Constrained Association Rules. In Proc. of 1998 ACMSIGMOD Intl. Conf. on Management of Data, pages 13–24, Seattle, WA, June 1998. [289] E. Omiecinski. Alternative Interest Measures for Mining Associations in Databases. IEEE Trans. on Knowledge and Data Engineering, 15(1):57–69, January/February 2003. [290] B. Ozden, S. Ramaswamy, and A. Silberschatz. Cyclic Association Rules. In Proc. of the 14th Intl. Conf. on Data Eng., pages 412–421, Orlando, FL, February 1998. [291] A. Ozgur, P. N. Tan, and V. Kumar. RBA: An Integrated Framework for Regression based on Association Rules. In Proc. of the SIAM Intl. Conf. on Data Mining, pages 210–221, Orlando, FL, April 2004.
400
Bibliography [292] J. S. Park, M.S. Chen, and P. S. Yu. An eﬀective hashbased algorithm for mining association rules. SIGMOD Record, 25(2):175–186, 1995. [293] S. Parthasarathy and M. Coatney. Eﬃcient Discovery of Common Substructures in Macromolecules. In Proc. of the 2002 IEEE Intl. Conf. on Data Mining, pages 362–369, Maebashi City, Japan, December 2002. [294] N. Pasquier, Y. Bastide, R. Taouil, and L. Lakhal. Discovering frequent closed itemsets for association rules. In Proc. of the 7th Intl. Conf. on Database Theory (ICDT’99), pages 398–416, Jerusalem, Israel, January 1999. [295] J. Pei, J. Han, H. J. Lu, S. Nishio, and S. Tang. HMine: HyperStructure Mining of Frequent Patterns in Large Databases. In Proc. of the 2001 IEEE Intl. Conf. on Data Mining, pages 441–448, San Jose, CA, November 2001. [296] J. Pei, J. Han, B. MortazaviAsl, and H. Zhu. Mining Access Patterns Eﬃciently from Web Logs. In Proc. of the 4th PacificAsia Conf. on Knowledge Discovery and Data Mining, pages 396–407, Kyoto, Japan, April 2000. [297] G. PiatetskyShapiro. Discovery, Analysis and Presentation of Strong Rules. In G. PiatetskyShapiro and W. Frawley, editors, Knowledge Discovery in Databases, pages 229–248. MIT Press, Cambridge, MA, 1991. [298] C. Potter, S. Klooster, M. Steinbach, P. N. Tan, V. Kumar, S. Shekhar, and C. Carvalho. Understanding Global Teleconnections of Climate to Regional Model Estimates of Amazon Ecosystem Carbon Fluxes. Global Change Biology, 10(5):693–703, 2004. [299] C. Potter, S. Klooster, M. Steinbach, P. N. Tan, V. Kumar, S. Shekhar, R. Myneni, and R. Nemani. Global Teleconnections of Ocean Climate to Terrestrial Carbon Flux. J. Geophysical Research, 108(D17), 2003. [300] G. D. Ramkumar, S. Ranka, and S. Tsur. Weighted Association Rules: Model and Algorithm. http://www.cs.ucla.edu/˜czdemo/tsur/, 1997. [301] S. Sarawagi, S. Thomas, and R. Agrawal. Integrating Mining with Relational Database Systems: Alternatives and Implications. In Proc. of 1998 ACMSIGMOD Intl. Conf. on Management of Data, pages 343–354, Seattle, WA, 1998. [302] K. Satou, G. Shibayama, T. Ono, Y. Yamamura, E. Furuichi, S. Kuhara, and T. Takagi. Finding Association Rules on Heterogeneous Genome Data. In Proc. of the Pacific Symp. on Biocomputing, pages 397–408, Hawaii, January 1997. [303] A. Savasere, E. Omiecinski, and S. Navathe. An eﬃcient algorithm for mining association rules in large databases. In Proc. of the 21st Int. Conf. on Very Large Databases (VLDB‘95), pages 432–444, Zurich, Switzerland, September 1995. [304] A. Savasere, E. Omiecinski, and S. Navathe. Mining for Strong Negative Associations in a Large Database of Customer Transactions. In Proc. of the 14th Intl. Conf. on Data Engineering, pages 494–502, Orlando, Florida, February 1998. [305] M. Seno and G. Karypis. LPMiner: An Algorithm for Finding Frequent Itemsets Using LengthDecreasing Support Constraint. In Proc. of the 2001 IEEE Intl. Conf. on Data Mining, pages 505–512, San Jose, CA, November 2001. [306] T. Shintani and M. Kitsuregawa. Hash based parallel algorithms for mining association rules. In Proc of the 4th Intl. Conf. on Parallel and Distributed Info. Systems, pages 19–30, Miami Beach, FL, December 1996. [307] A. Silberschatz and A. Tuzhilin. What makes patterns interesting in knowledge discovery systems. IEEE Trans. on Knowledge and Data Engineering, 8(6):970–974, 1996. [308] C. Silverstein, S. Brin, and R. Motwani. Beyond market baskets: Generalizing association rules to dependence rules. Data Mining and Knowledge Discovery, 2(1):39–68, 1998.
401
Chapter 6
Association Analysis
[309] E.H. Simpson. The Interpretation of Interaction in Contingency Tables. Journal of the Royal Statistical Society, B(13):238–241, 1951. [310] L. Singh, B. Chen, R. Haight, and P. Scheuermann. An Algorithm for Constrained Association Rule Mining in Semistructured Data. In Proc. of the 3rd PacificAsia Conf. on Knowledge Discovery and Data Mining, pages 148–158, Beijing, China, April 1999. [311] R. Srikant and R. Agrawal. Mining Quantitative Association Rules in Large Relational Tables. In Proc. of 1996 ACMSIGMOD Intl. Conf. on Management of Data, pages 1–12, Montreal, Canada, 1996. [312] R. Srikant and R. Agrawal. Mining Sequential Patterns: Generalizations and Performance Improvements. In Proc. of the 5th Intl Conf. on Extending Database Technology (EDBT’96), pages 18–32, Avignon, France, 1996. [313] R. Srikant, Q. Vu, and R. Agrawal. Mining Association Rules with Item Constraints. In Proc. of the 3rd Intl. Conf. on Knowledge Discovery and Data Mining, pages 67–73, Newport Beach, CA, August 1997. [314] M. Steinbach, P. N. Tan, and V. Kumar. Support Envelopes: A Technique for Exploring the Structure of Association Patterns. In Proc. of the 10th Intl. Conf. on Knowledge Discovery and Data Mining, pages 296–305, Seattle, WA, August 2004. [315] M. Steinbach, P. N. Tan, H. Xiong, and V. Kumar. Extending the Notion of Support. In Proc. of the 10th Intl. Conf. on Knowledge Discovery and Data Mining, pages 689– 694, Seattle, WA, August 2004. [316] E. Suzuki. Autonomous Discovery of Reliable Exception Rules. In Proc. of the 3rd Intl. Conf. on Knowledge Discovery and Data Mining, pages 259–262, Newport Beach, CA, August 1997. [317] P. N. Tan and V. Kumar. Mining Association Patterns in Web Usage Data. In Proc. of the Intl. Conf. on Advances in Infrastructure for eBusiness, eEducation, eScience and eMedicine on the Internet, L’Aquila, Italy, January 2002. [318] P. N. Tan, V. Kumar, and J. Srivastava. Selecting the Right Interestingness Measure for Association Patterns. In Proc. of the 8th Intl. Conf. on Knowledge Discovery and Data Mining, pages 32–41, Edmonton, Canada, July 2002. [319] P. N. Tan, M. Steinbach, V. Kumar, S. Klooster, C. Potter, and A. Torregrosa. Finding SpatioTemporal Patterns in Earth Science Data. In KDD 2001 Workshop on Temporal Data Mining, San Francisco, CA, 2001. [320] H. Toivonen. Sampling Large Databases for Association Rules. In Proc. of the 22nd VLDB Conf., pages 134–145, Bombay, India, 1996. [321] H. Toivonen, M. Klemettinen, P. Ronkainen, K. Hatonen, and H. Mannila. Pruning and Grouping Discovered Association Rules. In ECML95 Workshop on Statistics, Machine Learning and Knowledge Discovery in Databases, pages 47 – 52, Heraklion, Greece, April 1995. [322] S. Tsur, J. Ullman, S. Abiteboul, C. Clifton, R. Motwani, S. Nestorov, and A. Rosenthal. Query Flocks: A Generalization of Association Rule Mining. In Proc. of 1998 ACMSIGMOD Intl. Conf. on Management of Data, pages 1–12, Seattle, WA, June 1998. [323] A. Tung, H. J. Lu, J. Han, and L. Feng. Breaking the Barrier of Transactions: Mining InterTransaction Association Rules. In Proc. of the 5th Intl. Conf. on Knowledge Discovery and Data Mining, pages 297–301, San Diego, CA, August 1999. [324] K. Wang, Y. He, and J. Han. Mining Frequent Itemsets Using Support Constraints. In Proc. of the 26th VLDB Conf., pages 43–52, Cairo, Egypt, September 2000.
402
BIBLIOGRAPHY [325] K. Wang, S. H. Tay, and B. Liu. InterestingnessBased Interval Merger for Numeric Association Rules. In Proc. of the 4th Intl. Conf. on Knowledge Discovery and Data Mining, pages 121–128, New York, NY, August 1998. [326] G. I. Webb. Preliminary investigations into statistically valid exploratory rule discovery. In Proc. of the Australasian Data Mining Workshop (AusDM03), Canberra, Australia, December 2003. [327] H. Xiong, X. He, C. Ding, Y. Zhang, V. Kumar, and S. R. Holbrook. Identiﬁcation of Functional Modules in Protein Complexes via Hyperclique Pattern Discovery. In Proc. of the Pacific Symposium on Biocomputing, (PSB 2005), Maui, January 2005. [328] H. Xiong, S. Shekhar, P. N. Tan, and V. Kumar. Exploiting a Supportbased Upper Bound of Pearson’s Correlation Coeﬃcient for Eﬃciently Identifying Strongly Correlated Pairs. In Proc. of the 10th Intl. Conf. on Knowledge Discovery and Data Mining, pages 334–343, Seattle, WA, August 2004. [329] H. Xiong, M. Steinbach, P. N. Tan, and V. Kumar. HICAP: Hierarchial Clustering with Pattern Preservation. In Proc. of the SIAM Intl. Conf. on Data Mining, pages 279–290, Orlando, FL, April 2004. [330] H. Xiong, P. N. Tan, and V. Kumar. Mining Strong Aﬃnity Association Patterns in Data Sets with Skewed Support Distribution. In Proc. of the 2003 IEEE Intl. Conf. on Data Mining, pages 387–394, Melbourne, FL, 2003. [331] X. Yan and J. Han. gSpan: Graphbased Substructure Pattern Mining. In Proc. of the 2002 IEEE Intl. Conf. on Data Mining, pages 721–724, Maebashi City, Japan, December 2002. [332] C. Yang, U. M. Fayyad, and P. S. Bradley. Eﬃcient discovery of errortolerant frequent itemsets in high dimensions. In Proc. of the 7th Intl. Conf. on Knowledge Discovery and Data Mining, pages 194–203, San Francisco, CA, August 2001. [333] M. J. Zaki. Parallel and Distributed Association Mining: A Survey. IEEE Concurrency, special issue on Parallel Mechanisms for Data Mining, 7(4):14–25, December 1999. [334] M. J. Zaki. Generating NonRedundant Association Rules. In Proc. of the 6th Intl. Conf. on Knowledge Discovery and Data Mining, pages 34–43, Boston, MA, August 2000. [335] M. J. Zaki. Eﬃciently mining frequent trees in a forest. In Proc. of the 8th Intl. Conf. on Knowledge Discovery and Data Mining, pages 71–80, Edmonton, Canada, July 2002. [336] M. J. Zaki and M. Orihara. Theoretical foundations of association rules. In Proc. of the 1998 ACM SIGMOD Workshop on Research Issues in Data Mining and Knowledge Discovery, Seattle, WA, June 1998. [337] M. J. Zaki, S. Parthasarathy, M. Ogihara, and W. Li. New Algorithms for Fast Discovery of Association Rules. In Proc. of the 3rd Intl. Conf. on Knowledge Discovery and Data Mining, pages 283–286, Newport Beach, CA, August 1997. [338] H. Zhang, B. Padmanabhan, and A. Tuzhilin. On the Discovery of Signiﬁcant Statistical Quantitative Rules. In Proc. of the 10th Intl. Conf. on Knowledge Discovery and Data Mining, pages 374–383, Seattle, WA, August 2004. [339] Z. Zhang, Y. Lu, and B. Zhang. An Eﬀective PartioningCombining Algorithm for Discovering Quantitative Association Rules. In Proc. of the 1st PacificAsia Conf. on Knowledge Discovery and Data Mining, Singapore, 1997. [340] N. Zhong, Y. Y. Yao, and S. Ohsuga. Peculiarity Oriented Multidatabase Mining. In Proc. of the 3rd European Conf. of Principles and Practice of Knowledge Discovery in Databases, pages 136–146, Prague, Czech Republic, 1999.
403
Chapter 6
6.10
Association Analysis
Exercises
1. For each of the following questions, provide an example of an association rule from the market basket domain that satisﬁes the following conditions. Also, describe whether such rules are subjectively interesting. (a) A rule that has high support and high conﬁdence. (b) A rule that has reasonably high support but low conﬁdence. (c) A rule that has low support and low conﬁdence. (d) A rule that has low support and high conﬁdence. 2. Consider the data set shown in Table 6.22. Table 6.22. Example of market basket transactions. Customer ID 1 1 2 2 3 3 4 4 5 5
Transaction ID 0001 0024 0012 0031 0015 0022 0029 0040 0033 0038
Items Bought {a, d, e} {a, b, c, e} {a, b, d, e} {a, c, d, e} {b, c, e} {b, d, e} {c, d} {a, b, c} {a, d, e} {a, b, e}
(a) Compute the support for itemsets {e}, {b, d}, and {b, d, e} by treating each transaction ID as a market basket. (b) Use the results in part (a) to compute the conﬁdence for the association rules {b, d} −→ {e} and {e} −→ {b, d}. Is conﬁdence a symmetric measure? (c) Repeat part (a) by treating each customer ID as a market basket. Each item should be treated as a binary variable (1 if an item appears in at least one transaction bought by the customer, and 0 otherwise.) (d) Use the results in part (c) to compute the conﬁdence for the association rules {b, d} −→ {e} and {e} −→ {b, d}. (e) Suppose s1 and c1 are the support and conﬁdence values of an association rule r when treating each transaction ID as a market basket. Also, let s2 and c2 be the support and conﬁdence values of r when treating each customer ID as a market basket. Discuss whether there are any relationships between s1 and s2 or c1 and c2 .
404
6.10 3.
Exercises
(a) What is the conﬁdence for the rules ∅ −→ A and A −→ ∅? (b) Let c1 , c2 , and c3 be the conﬁdence values of the rules {p} −→ {q}, {p} −→ {q, r}, and {p, r} −→ {q}, respectively. If we assume that c1 , c2 , and c3 have diﬀerent values, what are the possible relationships that may exist among c1 , c2 , and c3 ? Which rule has the lowest conﬁdence? (c) Repeat the analysis in part (b) assuming that the rules have identical support. Which rule has the highest conﬁdence? (d) Transitivity: Suppose the conﬁdence of the rules A −→ B and B −→ C are larger than some threshold, minconf . Is it possible that A −→ C has a conﬁdence less than minconf ?
4. For each of the following measures, determine whether it is monotone, antimonotone, or nonmonotone (i.e., neither monotone nor antimonotone). Example: Support, s = s(Y ) whenever X ⊂ Y .
σ(X) T 
is antimonotone because s(X) ≥
(a) A characteristic rule is a rule of the form {p} −→ {q1 , q2 , . . . , qn }, where the rule antecedent contains only a single item. An itemset of size k can produce up to k characteristic rules. Let ζ be the minimum conﬁdence of all characteristic rules generated from a given itemset: ζ({p1 , p2 , . . . , pk }) = min c {p1 } −→ {p2 , p3 , . . . , pk } , . . . c {pk } −→ {p1 , p3 . . . , pk−1 } Is ζ monotone, antimonotone, or nonmonotone? (b) A discriminant rule is a rule of the form {p1 , p2 , . . . , pn } −→ {q}, where the rule consequent contains only a single item. An itemset of size k can produce up to k discriminant rules. Let η be the minimum conﬁdence of all discriminant rules generated from a given itemset: η({p1 , p2 , . . . , pk })
= min c {p2 , p3 , . . . , pk } −→ {p1 } , . . . c {p1 , p2 , . . . pk−1 } −→ {pk }
Is η monotone, antimonotone, or nonmonotone? (c) Repeat the analysis in parts (a) and (b) by replacing the min function with a max function. 5. Prove Equation 6.3. (Hint: First, count the number of ways to create an itemset that forms the left hand side of the rule. Next, for each size k itemset selected for the lefthand side, count the number of ways to choose the remaining d − k items to form the righthand side of the rule.)
405
Chapter 6
Association Analysis
Table 6.23. Market basket transactions. Transaction ID Items Bought 1 {Milk, Beer, Diapers} 2 {Bread, Butter, Milk} 3 {Milk, Diapers, Cookies} 4 {Bread, Butter, Cookies} 5 {Beer, Cookies, Diapers} 6 {Milk, Diapers, Bread, Butter} 7 {Bread, Butter, Diapers} 8 {Beer, Diapers} 9 {Milk, Diapers, Bread, Butter} 10 {Beer, Cookies} 6. Consider the market basket transactions shown in Table 6.23. (a) What is the maximum number of association rules that can be extracted from this data (including rules that have zero support)? (b) What is the maximum size of frequent itemsets that can be extracted (assuming minsup > 0)? (c) Write an expression for the maximum number of size3 itemsets that can be derived from this data set. (d) Find an itemset (of size 2 or larger) that has the largest support. (e) Find a pair of items, a and b, such that the rules {a} −→ {b} and {b} −→ {a} have the same conﬁdence. 7. Consider the following set of frequent 3itemsets: {1, 2, 3}, {1, 2, 4}, {1, 2, 5}, {1, 3, 4}, {1, 3, 5}, {2, 3, 4}, {2, 3, 5}, {3, 4, 5}. Assume that there are only ﬁve items in the data set. (a) List all candidate 4itemsets obtained by a candidate generation procedure using the Fk−1 × F1 merging strategy. (b) List all candidate 4itemsets obtained by the candidate generation procedure in Apriori. (c) List all candidate 4itemsets that survive the candidate pruning step of the Apriori algorithm. 8. The Apriori algorithm uses a generateandcount strategy for deriving frequent itemsets. Candidate itemsets of size k + 1 are created by joining a pair of frequent itemsets of size k (this is known as the candidate generation step). A candidate is discarded if any one of its subsets is found to be infrequent during the candidate pruning step. Suppose the Apriori algorithm is applied to the
406
6.10
Exercises
Table 6.24. Example of market basket transactions. Transaction ID 1 2 3 4 5 6 7 8 9 10
Items Bought {a, b, d, e} {b, c, d} {a, b, d, e} {a, c, d, e} {b, c, d, e} {b, d, e} {c, d} {a, b, c} {a, d, e} {b, d}
data set shown in Table 6.24 with minsup = 30%, i.e., any itemset occurring in less than 3 transactions is considered to be infrequent. (a) Draw an itemset lattice representing the data set given in Table 6.24. Label each node in the lattice with the following letter(s): • N: If the itemset is not considered to be a candidate itemset by the Apriori algorithm. There are two reasons for an itemset not to be considered as a candidate itemset: (1) it is not generated at all during the candidate generation step, or (2) it is generated during the candidate generation step but is subsequently removed during the candidate pruning step because one of its subsets is found to be infrequent. • F: If the candidate itemset is found to be frequent by the Apriori algorithm. • I: If the candidate itemset is found to be infrequent after support counting. (b) What is the percentage of frequent itemsets (with respect to all itemsets in the lattice)? (c) What is the pruning ratio of the Apriori algorithm on this data set? (Pruning ratio is deﬁned as the percentage of itemsets not considered to be a candidate because (1) they are not generated during candidate generation or (2) they are pruned during the candidate pruning step.) (d) What is the false alarm rate (i.e, percentage of candidate itemsets that are found to be infrequent after performing support counting)? 9. The Apriori algorithm uses a hash tree data structure to eﬃciently count the support of candidate itemsets. Consider the hash tree for candidate 3itemsets shown in Figure 6.32.
407
Chapter 6
Association Analysis 1,4,7
1,4,7 L1 {145} {178}
2,5,8
L3 {125} {158} {458}
1,4,7
3,6,9
3,6,9 2,5,8
L5 {168}
1,4,7 2,5,8 L2 {127} {457}
3,6,9
2,5,8
3,6,9 L4 {459} {456} {789}
L6 {246} {278}
L7 {258} {289}
L8 {568}
1,4,7 3,6,9 2,5,8 L9 L11 L12 {346} {356} {367} {379} {689} {678}
Figure 6.32. An example of a hash tree structure.
(a) Given a transaction that contains items {1, 3, 4, 5, 8}, which of the hash tree leaf nodes will be visited when ﬁnding the candidates of the transaction? (b) Use the visited leaf nodes in part (b) to determine the candidate itemsets that are contained in the transaction {1, 3, 4, 5, 8}. 10. Consider the following set of candidate 3itemsets: {1, 2, 3}, {1, 2, 6}, {1, 3, 4}, {2, 3, 4}, {2, 4, 5}, {3, 4, 6}, {4, 5, 6} (a) Construct a hash tree for the above candidate 3itemsets. Assume the tree uses a hash function where all oddnumbered items are hashed to the left child of a node, while the evennumbered items are hashed to the right child. A candidate kitemset is inserted into the tree by hashing on each successive item in the candidate and then following the appropriate branch of the tree according to the hash value. Once a leaf node is reached, the candidate is inserted based on one of the following conditions: Condition 1: If the depth of the leaf node is equal to k (the root is assumed to be at depth 0), then the candidate is inserted regardless of the number of itemsets already stored at the node. Condition 2: If the depth of the leaf node is less than k, then the candidate can be inserted as long as the number of itemsets stored at the node is less than maxsize. Assume maxsize = 2 for this question. Condition 3: If the depth of the leaf node is less than k and the number of itemsets stored at the node is equal to maxsize, then the leaf node is converted into an internal node. New leaf nodes are created as children of the old leaf node. Candidate itemsets previously stored
408
6.10
Exercises
null
a
b
c
d
e
ab
ac
ad
ae
bc
bd
be
cd
ce
de
abc
abd
abe
acd
ace
ade
bcd
bce
bde
cde
abcd
abce
abde
acde
bcde
abcde
Figure 6.33. An itemset lattice
in the old leaf node are distributed to the children based on their hash values. The new candidate is also hashed to its appropriate leaf node. (b) How many leaf nodes are there in the candidate hash tree? How many internal nodes are there? (c) Consider a transaction that contains the following items: {1, 2, 3, 5, 6}. Using the hash tree constructed in part (a), which leaf nodes will be checked against the transaction? What are the candidate 3itemsets contained in the transaction? 11. Given the lattice structure shown in Figure 6.33 and the transactions given in Table 6.24, label each node with the following letter(s): • M if the node is a maximal frequent itemset, • C if it is a closed frequent itemset, • N if it is frequent but neither maximal nor closed, and • I if it is infrequent. Assume that the support threshold is equal to 30%. 12. The original association rule mining formulation uses the support and conﬁdence measures to prune uninteresting rules.
409
Chapter 6
Association Analysis
(a) Draw a contingency table for each of the following rules using the transactions shown in Table 6.25. Table 6.25. Example of market basket transactions. Transaction ID 1 2 3 4 5 6 7 8 9 10
Items Bought {a, b, d, e} {b, c, d} {a, b, d, e} {a, c, d, e} {b, c, d, e} {b, d, e} {c, d} {a, b, c} {a, d, e} {b, d}
Rules: {b} −→ {c}, {a} −→ {d}, {b} −→ {d}, {e} −→ {c}, {c} −→ {a}. (b) Use the contingency tables in part (a) to compute and rank the rules in decreasing order according to the following measures. i. Support. ii. Conﬁdence. iii. Interest(X −→ Y ) = iv. IS(X −→ Y ) = √
P (X,Y ) P (X) P (Y P (X,Y )
P (X)P (Y )
v. Klosgen(X −→ Y ) = P (X,Y ) P (X) .
).
.
P (X, Y )×(P (Y X)−P (Y )), where P (Y X) =
vi. Odds ratio(X −→ Y ) =
P (X,Y )P (X,Y ) . P (X,Y )P (X,Y )
13. Given the rankings you had obtained in Exercise 12, compute the correlation between the rankings of conﬁdence and the other ﬁve measures. Which measure is most highly correlated with conﬁdence? Which measure is least correlated with conﬁdence? 14. Answer the following questions using the data sets shown in Figure 6.34. Note that each data set contains 1000 items and 10,000 transactions. Dark cells indicate the presence of items and white cells indicate the absence of items. We will apply the Apriori algorithm to extract frequent itemsets with minsup = 10% (i.e., itemsets must be contained in at least 1000 transactions)? (a) Which data set(s) will produce the most number of frequent itemsets?
410
6.10
Exercises
(b) Which data set(s) will produce the fewest number of frequent itemsets? (c) Which data set(s) will produce the longest frequent itemset? (d) Which data set(s) will produce frequent itemsets with highest maximum support? (e) Which data set(s) will produce frequent itemsets containing items with widevarying support levels (i.e., items with mixed support, ranging from less than 20% to more than 70%). 15.
(a) Prove that the φ coeﬃcient is equal to 1 if and only if f11 = f1+ = f+1 . (b) Show that if A and B are independent, then P (A, B)×P (A, B) = P (A, B)× P (A, B). (c) Show that Yule’s Q and Y coeﬃcients f11 f00 − f10 f01 Q = f11 f00 + f10 f01 √ √ f11 f00 − f10 f01 √ √ Y = f11 f00 + f10 f01 are normalized versions of the odds ratio. (d) Write a simpliﬁed expression for the value of each measure shown in Tables 6.11 and 6.12 when the variables are statistically independent.
16. Consider the interestingness measure, M = rule A −→ B.
P (BA)−P (B) , 1−P (B)
for an association
(a) What is the range of this measure? When does the measure attain its maximum and minimum values? (b) How does M behave when P (A, B) is increased while P (A) and P (B) remain unchanged? (c) How does M behave when P (A) is increased while P (A, B) and P (B) remain unchanged? (d) How does M behave when P (B) is increased while P (A, B) and P (A) remain unchanged? (e) Is the measure symmetric under variable permutation? (f) What is the value of the measure when A and B are statistically independent? (g) Is the measure nullinvariant? (h) Does the measure remain invariant under row or column scaling operations? (i) How does the measure behave under the inversion operation?
411
Chapter 6
Association Analysis
Items
Items
4000 6000
2000 Transactions
Transactions
2000
4000 6000
8000
8000
200 400 600 800 (a)
200 400 600 800 (b)
Items
Items
2000
4000 6000
Transactions
Transactions
2000
4000 6000 8000
8000
200 400 600 800 (c)
200 400 600 800 (d)
Items
Items
2000
4000 6000
Transactions
Transactions
2000 10% are 1s 90% are 0s (uniformly distributed)
200 400 600 800 (f)
Figure 6.34. Figures for Exercise 14.
412
6000 8000
8000
200 400 600 800 (e)
4000
6.10
Exercises
17. Suppose we have market basket data consisting of 100 transactions and 20 items. If the support for item a is 25%, the support for item b is 90% and the support for itemset {a, b} is 20%. Let the support and conﬁdence thresholds be 10% and 60%, respectively. (a) Compute the conﬁdence of the association rule {a} → {b}. Is the rule interesting according to the conﬁdence measure? (b) Compute the interest measure for the association pattern {a, b}. Describe the nature of the relationship between item a and item b in terms of the interest measure. (c) What conclusions can you draw from the results of parts (a) and (b)? (d) Prove that if the conﬁdence of the rule {a} −→ {b} is less than the support of {b}, then: i. c({a} −→ {b}) > c({a} −→ {b}), ii. c({a} −→ {b}) > s({b}), where c(·) denote the rule conﬁdence and s(·) denote the support of an itemset. 18. Table 6.26 shows a 2 × 2 × 2 contingency table for the binary variables A and B at diﬀerent values of the control variable C. Table 6.26. A Contingency Table. A
C=0
B
C=1
B
1
0
1
0
15
0
15
30
1
5
0
0
0
15
(a) Compute the φ coeﬃcient for A and B when C = 0, C = 1, and C = 0 or . 1. Note that φ({A, B}) = √ P (A,B)−P (A)P (B) P (A)P (B)(1−P (A))(1−P (B))
(b) What conclusions can you draw from the above result? 19. Consider the contingency tables shown in Table 6.27. (a) For table I, compute support, the interest measure, and the φ correlation coeﬃcient for the association pattern {A, B}. Also, compute the conﬁdence of rules A → B and B → A.
413
Chapter 6
Association Analysis Table 6.27. Contingency tables for Exercise 19. B
B
A
9
1
A
1
89
(a) Table I.
B
B
A
89
1
A
1
9
(b) Table II.
(b) For table II, compute support, the interest measure, and the φ correlation coeﬃcient for the association pattern {A, B}. Also, compute the conﬁdence of rules A → B and B → A. (c) What conclusions can you draw from the results of (a) and (b)? 20. Consider the relationship between customers who buy highdeﬁnition televisions and exercise machines as shown in Tables 6.19 and 6.20. (a) Compute the odds ratios for both tables. (b) Compute the φcoeﬃcient for both tables. (c) Compute the interest factor for both tables. For each of the measures given above, describe how the direction of association changes when data is pooled together instead of being stratiﬁed.
414
7 Association Analysis: Advanced Concepts The association rule mining formulation described in the previous chapter assumes that the input data consists of binary attributes called items. The presence of an item in a transaction is also assumed to be more important than its absence. As a result, an item is treated as an asymmetric binary attribute and only frequent patterns are considered interesting. This chapter extends the formulation to data sets with symmetric binary, categorical, and continuous attributes. The formulation will also be extended to incorporate more complex entities such as sequences and graphs. Although the overall structure of association analysis algorithms remains unchanged, certain aspects of the algorithms must be modiﬁed to handle the nontraditional entities.
7.1
Handling Categorical Attributes
There are many applications that contain symmetric binary and nominal attributes. The Internet survey data shown in Table 7.1 contains symmetric binary attributes such as Gender, Computer at Home, Chat Online, Shop Online, and Privacy Concerns; as well as nominal attributes such as Level of Education and State. Using association analysis, we may uncover interesting information about the characteristics of Internet users such as: {Shop Online = Yes} −→ {Privacy Concerns = Yes}. This rule suggests that most Internet users who shop online are concerned about their personal privacy.
From Chapter 7 of Introduction to Data Mining, First Edition. PangNing Tan, Michael Steinbach, Vipin Kumar. Copyright © 2006 by Pearson Education, Inc. All rights reserved.
415
Chapter 7
Association Analysis: Advanced Concepts
Table 7.1. Internet survey data with categorical attributes. Gender Female Male Male Female Female Male Male Male Female ...
Level of Education Graduate College Graduate College Graduate College College High School Graduate ...
State Illinois California Michigan Virginia California Minnesota Alaska Oregon Texas ...
Computer at Home Yes No Yes No Yes Yes Yes Yes No ...
Chat Online Yes No Yes No No Yes Yes No Yes ...
Shop Online Yes No Yes Yes No Yes Yes No No ...
Privacy Concerns Yes No Yes Yes Yes Yes No No No ...
To extract such patterns, the categorical and symmetric binary attributes are transformed into “items” ﬁrst, so that existing association rule mining algorithms can be applied. This type of transformation can be performed by creating a new item for each distinct attributevalue pair. For example, the nominal attribute Level of Education can be replaced by three binary items: Education = College, Education = Graduate, and Education = High School. Similarly, symmetric binary attributes such as Gender can be converted into a pair of binary items, Male and Female. Table 7.2 shows the result of binarizing the Internet survey data. Table 7.2. Internet survey data after binarizing categorical and symmetric binary attributes.
416
Male
Female
0 1 1 0 0 1 1 1 0 ...
1 0 0 1 1 0 0 0 1 ...
Education = Graduate 1 0 1 0 1 0 0 0 1 ...
Education = College 0 1 0 1 0 1 1 0 0 ...
... ... ... ... ... ... ... ... ... ... ...
Privacy = Yes 1 0 1 1 1 1 0 0 0 ...
Privacy = No 0 1 0 0 0 0 1 1 1 ...
7.1
Handling Categorical Attributes
There are several issues to consider when applying association analysis to the binarized data: 1. Some attribute values may not be frequent enough to be part of a frequent pattern. This problem is more evident for nominal attributes that have many possible values, e.g., state names. Lowering the support threshold does not help because it exponentially increases the number of frequent patterns found (many of which may be spurious) and makes the computation more expensive. A more practical solution is to group related attribute values into a small number of categories. For example, each state name can be replaced by its corresponding geographical region, such as Midwest, Pacific Northwest, Southwest, and East Coast. Another possibility is to aggregate the less frequent attribute values into a single category called Others, as shown in Figure 7.1. Virginia Others Ohio New York
Illinois Michigan Florida
California Minnesota Texas
Massachusetts Oregon
Figure 7.1. A pie chart with a merged category called Others.
2. Some attribute values may have considerably higher frequencies than others. For example, suppose 85% of the survey participants own a home computer. By creating a binary item for each attribute value that appears frequently in the data, we may potentially generate many redundant patterns, as illustrated by the following example: {Computer at home = Yes, Shop Online = Yes} −→ {Privacy Concerns = Yes}.
417
Chapter 7
Association Analysis: Advanced Concepts
The rule is redundant because it is subsumed by the more general rule given at the beginning of this section. Because the highfrequency items correspond to the typical values of an attribute, they seldom carry any new information that can help us to better understand the pattern. It may therefore be useful to remove such items before applying standard association analysis algorithms. Another possibility is to apply the techniques presented in Section 6.8 for handling data sets with a wide range of support values. 3. Although the width of every transaction is the same as the number of attributes in the original data, the computation time may increase especially when many of the newly created items become frequent. This is because more time is needed to deal with the additional candidate itemsets generated by these items (see Exercise 1 on page 473). One way to reduce the computation time is to avoid generating candidate itemsets that contain more than one item from the same attribute. For example, we do not have to generate a candidate itemset such as {State = X, State = Y, . . .} because the support count of the itemset is zero.
7.2
Handling Continuous Attributes
The Internet survey data described in the previous section may also contain continuous attributes such as the ones shown in Table 7.3. Mining the continuous attributes may reveal useful insights about the data such as “users whose annual income is more than $120K belong to the 45–60 age group” or “users who have more than 3 email accounts and spend more than 15 hours online per week are often concerned about their personal privacy.” Association rules that contain continuous attributes are commonly known as quantitative association rules. This section describes the various methodologies for applying association analysis to continuous data. We will speciﬁcally discuss three types of methods: (1) discretizationbased methods, (2) statisticsbased methods, and (3) nondiscretization methods. The quantitative association rules derived using these methods are quite diﬀerent in nature.
7.2.1
DiscretizationBased Methods
Discretization is the most common approach for handling continuous attributes. This approach groups the adjacent values of a continuous attribute into a ﬁnite number of intervals. For example, the Age attribute can be divided into the
418
7.2
Handling Continuous Attributes
Table 7.3. Internet survey data with continuous attributes. Gender
...
Age
Female Male Male Female Female Male Male Male Female ...
... ... ... ... ... ... ... ... ... ...
26 51 29 45 31 25 37 41 26 ...
Annual Income 90K 135K 80K 120K 95K 55K 100K 65K 85K ...
No. of Hours Spent Online per Week 20 10 10 15 20 25 10 8 12 ...
No. of Email Accounts 4 2 3 3 5 5 1 2 1 ...
Privacy Concern Yes No Yes Yes Yes Yes No No No ...
following intervals: Age ∈ [12, 16), Age ∈ [16, 20), Age ∈ [20, 24), . . . , Age ∈ [56, 60), where [a, b) represents an interval that includes a but not b. Discretization can be performed using any of the techniques described in Section 2.3.6 (equal interval width, equal frequency, entropybased, or clustering). The discrete intervals are then mapped into asymmetric binary attributes so that existing association analysis algorithms can be applied. Table 7.4 shows the Internet survey data after discretization and binarization. Table 7.4. Internet survey data after binarizing categorical and continuous attributes. Male
Female
...
0 1 1 0 0 1 1 1 0 ...
1 0 0 1 1 0 0 0 1 ...
... ... ... ... ... ... ... ... ... ...
Age < 13 0 0 0 0 0 0 0 0 0 ...
Age ∈ [13, 21) 0 0 0 0 0 0 0 0 0 ...
Age ∈ [21, 30) 1 0 1 0 0 1 0 0 1 ...
... ... ... ... ... ... ... ... ... ... ...
Privacy = Yes 1 0 1 1 1 1 0 0 0 ...
Privacy = No 0 1 0 0 0 0 1 1 1 ...
419
Chapter 7
Association Analysis: Advanced Concepts
Table 7.5. A breakdown of Internet users who participated in online chat according to their age group. Age Group [12, 16) [16, 20) [20, 24) [24, 28) [28, 32) [32, 36) [36, 40) [40, 44) [44, 48) [48, 52) [52, 56) [56, 60)
Chat Online = Yes 12 11 11 12 14 15 16 16 4 5 5 4
Chat Online = No 13 2 3 13 12 12 14 14 10 11 10 11
A key parameter in attribute discretization is the number of intervals used to partition each attribute. This parameter is typically provided by the users and can be expressed in terms of the interval width (for the equal interval width approach), the average number of transactions per interval (for the equal frequency approach), or the number of desired clusters (for the clusteringbased approach). The diﬃculty in determining the right number of intervals can be illustrated using the data set shown in Table 7.5, which summarizes the responses of 250 users who participated in the survey. There are two strong rules embedded in the data: R1 : R2 :
Age ∈ [16, 24) −→ Chat Online = Yes (s = 8.8%, c = 81.5%). Age ∈ [44, 60) −→ Chat Online = No (s = 16.8%, c = 70%).
These rules suggest that most of the users from the age group of 16–24 often participate in online chatting, while those from the age group of 44–60 are less likely to chat online. In this example, we consider a rule to be interesting only if its support (s) exceeds 5% and its conﬁdence (c) exceeds 65%. One of the problems encountered when discretizing the Age attribute is how to determine the interval width. 1. If the interval is too wide, then we may lose some patterns because of their lack of conﬁdence. For example, when the interval width is 24 years, R1 and R2 are replaced by the following rules: R1 : R2 :
420
Age ∈ [12, 36) −→ Chat Online = Yes (s = 30%, c = 57.7%). Age ∈ [36, 60) −→ Chat Online = No (s = 28%, c = 58.3%).
7.2
Handling Continuous Attributes
Despite their higher supports, the wider intervals have caused the conﬁdence for both rules to drop below the minimum conﬁdence threshold. As a result, both patterns are lost after discretization. 2. If the interval is too narrow, then we may lose some patterns because of their lack of support. For example, if the interval width is 4 years, then R1 is broken up into the following two subrules: (4)
R11 : (4) R12 :
Age ∈ [16, 20) −→ Chat Online = Yes (s=4.4%, c=84.6%). Age ∈ [20, 24) −→ Chat Online = No (s=4.4%, c=78.6%).
Since the supports for the subrules are less than the minimum support threshold, R1 is lost after discretization. Similarly, the rule R2 , which is broken up into four subrules, will also be lost because the support of each subrule is less than the minimum support threshold. 3. If the interval width is 8 years, then the rule R2 is broken up into the following two subrules: (8)
R21 : (8) R22 :
Age ∈ [44, 52) −→ Chat Online = No (s=8.4%, c=70%). Age ∈ [52, 60) −→ Chat Online = No (s=8.4%, c=70%). (8)
(8)
Since R21 and R22 have suﬃcient support and conﬁdence, R2 can be recovered by aggregating both subrules. Meanwhile, R1 is broken up into the following two subrules: (8)
R11 : (8) R12 :
Age ∈ [12, 20) −→ Chat Online = Yes (s=9.2%, c=60.5%). Age ∈ [20, 28) −→ Chat Online = Yes (s=9.2%, c=60.0%).
Unlike R2 , we cannot recover the rule R1 by aggregating the subrules because both subrules fail the conﬁdence threshold. One way to address these issues is to consider every possible grouping of adjacent intervals. For example, we can start with an interval width of 4 years and then merge the adjacent intervals into wider intervals, Age ∈ [12, 16), Age ∈ [12, 20), . . . , Age ∈ [12, 60), Age ∈ [16, 20), Age ∈ [16, 24), etc. This approach enables the detection of both R1 and R2 as strong rules. However, it also leads to the following computational issues: 1. The computation becomes extremely expensive. If the range is initially divided into k intervals, then k(k − 1)/2 binary items must be
421
Chapter 7
Association Analysis: Advanced Concepts
generated to represent all possible intervals. Furthermore, if an item corresponding to the interval [a,b) is frequent, then all other items corresponding to intervals that subsume [a,b) must be frequent too. This approach can therefore generate far too many candidate and frequent itemsets. To address these problems, a maximum support threshold can be applied to prevent the creation of items corresponding to very wide intervals and to reduce the number of itemsets. 2. Many redundant rules are extracted. For example, consider the following pair of rules: R3 : {Age ∈ [16, 20), Gender = Male} −→ {Chat Online = Yes}, R4 : {Age ∈ [16, 24), Gender = Male} −→ {Chat Online = Yes}. R4 is a generalization of R3 (and R3 is a specialization of R4 ) because R4 has a wider interval for the Age attribute. If the conﬁdence values for both rules are the same, then R4 should be more interesting because it covers more examples—including those for R3 . R3 is therefore a redundant rule.
7.2.2
StatisticsBased Methods
Quantitative association rules can be used to infer the statistical properties of a population. For example, suppose we are interested in ﬁnding the average age of certain groups of Internet users based on the data provided in Tables 7.1 and 7.3. Using the statisticsbased method described in this section, quantitative association rules such as the following can be extracted: {Annual Income > $100K, Shop Online = Yes} −→ Age: Mean = 38. The rule states that the average age of Internet users whose annual income exceeds $100K and who shop online regularly is 38 years old. Rule Generation To generate the statisticsbased quantitative association rules, the target attribute used to characterize interesting segments of the population must be speciﬁed. By withholding the target attribute, the remaining categorical and continuous attributes in the data are binarized using the methods described in the previous section. Existing algorithms such as Apriori or FPgrowth are then applied to extract frequent itemsets from the binarized data. Each
422
7.2
Handling Continuous Attributes
frequent itemset identiﬁes an interesting segment of the population. The distribution of the target attribute in each segment can be summarized using descriptive statistics such as mean, median, variance, or absolute deviation. For example, the preceding rule is obtained by averaging the age of Internet users who support the frequent itemset {Annual Income > $100K, Shop Online = Yes}. The number of quantitative association rules discovered using this method is the same as the number of extracted frequent itemsets. Because of the way the quantitative association rules are deﬁned, the notion of conﬁdence is not applicable to such rules. An alternative method for validating the quantitative association rules is presented next. Rule Validation A quantitative association rule is interesting only if the statistics computed from transactions covered by the rule are diﬀerent than those computed from transactions not covered by the rule. For example, the rule given at the beginning of this section is interesting only if the average age of Internet users who do not support the frequent itemset {Annual Income > 100K, Shop Online = Yes} is signiﬁcantly higher or lower than 38 years old. To determine whether the diﬀerence in their average ages is statistically signiﬁcant, statistical hypothesis testing methods should be applied. Consider the quantitative association rule, A −→ t : µ, where A is a frequent itemset, t is the continuous target attribute, and µ is the average value of t among transactions covered by A. Furthermore, let µ denote the average value of t among transactions not covered by A. The goal is to test whether the diﬀerence between µ and µ is greater than some userspeciﬁed threshold, ∆. In statistical hypothesis testing, two opposite propositions, known as the null hypothesis and the alternative hypothesis, are given. A hypothesis test is performed to determine which of these two hypotheses should be accepted, based on evidence gathered from the data (see Appendix C). In this case, assuming that µ < µ , the null hypothesis is H0 : µ = µ + ∆, while the alternative hypothesis is H1 : µ > µ + ∆. To determine which hypothesis should be accepted, the following Zstatistic is computed: µ − µ − ∆ Z= 2 , s1 s22 + n1 n2
(7.1)
where n1 is the number of transactions supporting A, n2 is the number of transactions not supporting A, s1 is the standard deviation for t among transactions
423
Chapter 7
Association Analysis: Advanced Concepts
that support A, and s2 is the standard deviation for t among transactions that do not support A. Under the null hypothesis, Z has a standard normal distribution with mean 0 and variance 1. The value of Z computed using Equation 7.1 is then compared against a critical value, Zα , which is a threshold that depends on the desired conﬁdence level. If Z > Zα , then the null hypothesis is rejected and we may conclude that the quantitative association rule is interesting. Otherwise, there is not enough evidence in the data to show that the diﬀerence in mean is statistically signiﬁcant. Example 7.1. Consider the quantitative association rule {Income > 100K, Shop Online = Y es} −→ Age : µ = 38. Suppose there are 50 Internet users who supported the rule antecedent. The standard deviation of their ages is 3.5. On the other hand, the average age of the 200 users who do not support the rule antecedent is 30 and their standard deviation is 6.5. Assume that a quantitative association rule is considered interesting only if the diﬀerence between µ and µ is more than 5 years. Using Equation 7.1 we obtain 38 − 30 − 5 = 4.4414. Z=
3.52 6.52 + 50 200 For a onesided hypothesis test at a 95% conﬁdence level, the critical value for rejecting the null hypothesis is 1.64. Since Z > 1.64, the null hypothesis can be rejected. We therefore conclude that the quantitative association rule is interesting because the diﬀerence between the average ages of users who support and do not support the rule antecedent is more than 5 years.
7.2.3
Nondiscretization Methods
There are certain applications in which analysts are more interested in ﬁnding associations among the continuous attributes, rather than associations among discrete intervals of the continuous attributes. For example, consider the problem of ﬁnding word associations in text documents, as shown in Table 7.6. Each entry in the documentword matrix represents the normalized frequency count of a word appearing in a given document. The data is normalized by dividing the frequency of each word by the sum of the word frequency across all documents. One reason for this normalization is to make sure that the resulting support value is a number between 0 and 1. However, a more
424
7.2
Document d1 d2 d3 d4 d5
Handling Continuous Attributes
Table 7.6. Normalized documentword matrix. word1 word2 word3 word4 word5 0.3 0.6 0 0 0 0.1 0.2 0 0 0 0.4 0.2 0.7 0 0 0.2 0 0.3 0 0 0 0 0 1.0 1.0
word6 0.2 0.2 0.2 0.1 0.3
important reason is to ensure that the data is on the same scale so that sets of words that vary in the same way have similar support values. In text mining, analysts are more interested in ﬁnding associations between words (e.g., data and mining) instead of associations between ranges of word frequencies (e.g., data ∈ [1, 4] and mining ∈ [2, 3]). One way to do this is to transform the data into a 0/1 matrix, where the entry is 1 if the normalized frequency count exceeds some threshold t, and 0 otherwise. While this approach allows analysts to apply existing frequent itemset generation algorithms to the binarized data set, ﬁnding the right threshold for binarization can be quite tricky. If the threshold is set too high, it is possible to miss some interesting associations. Conversely, if the threshold is set too low, there is a potential for generating a large number of spurious associations. This section presents another methodology for ﬁnding word associations known as minApriori. Analogous to traditional association analysis, an itemset is considered to be a collection of words, while its support measures the degree of association among the words. The support of an itemset can be computed based on the normalized frequency of its corresponding words. For example, consider the document d1 shown in Table 7.6. The normalized frequencies for word1 and word2 in this document are 0.3 and 0.6, respectively. One might think that a reasonable approach to compute the association between both words is to take the average value of their normalized frequencies, i.e., (0.3 + 0.6)/2 = 0.45. The support of an itemset can then be computed by summing up the averaged normalized frequencies across all the documents: s({word1 , word2 }) =
0.3 + 0.6 0.1 + 0.2 0.4 + 0.2 0.2 + 0 + + + = 1. 2 2 2 2
This result is by no means an accident. Because every word frequency is normalized to 1, averaging the normalized frequencies makes the support for every itemset equal to 1. All itemsets are therefore frequent using this approach, making it useless for identifying interesting patterns.
425
Chapter 7
Association Analysis: Advanced Concepts
In minApriori, the association among words in a given document is obtained by taking the minimum value of their normalized frequencies, i.e., min(word1 , word2 ) = min(0.3, 0.6) = 0.3. The support of an itemset is computed by aggregating its association over all the documents. s({word1 , word2 }) = min(0.3, 0.6) + min(0.1, 0.2) + min(0.4, 0.2) + min(0.2, 0) = 0.6. The support measure deﬁned in minApriori has the following desired properties, which makes it suitable for ﬁnding word associations in documents: 1. Support increases monotonically as the normalized frequency of a word increases. 2. Support increases monotonically as the number of documents that contain the word increases. 3. Support has an antimonotone property. For example, consider a pair of itemsets {A, B} and {A, B, C}. Since min({A, B}) ≥ min({A, B, C}), s({A, B}) ≥ s({A, B, C}). Therefore, support decreases monotonically as the number of words in an itemset increases. The standard Apriori algorithm can be modiﬁed to ﬁnd associations among words using the new support deﬁnition.
7.3
Handling a Concept Hierarchy
A concept hierarchy is a multilevel organization of the various entities or concepts deﬁned in a particular domain. For example, in market basket analysis, a concept hierarchy has the form of an item taxonomy describing the “isa” relationships among items sold at a grocery store—e.g., milk is a kind of food and DVD is a kind of home electronics equipment (see Figure 7.2). Concept hierarchies are often deﬁned according to domain knowledge or based on a standard classiﬁcation scheme deﬁned by certain organizations (e.g., the Library of Congress classiﬁcation scheme is used to organize library materials based on their subject categories). A concept hierarchy can be represented using a directed acyclic graph, as shown in Figure 7.2. If there is an edge in the graph from a node p to another node c, we call p the parent of c and c the child of p. For example,
426
7.3
Handling a Concept Hierarchy
Food
Electronics
Bread Milk
Computers
Home Laptop Accessories
Wheat
White
Skim
2%
Desktop
TV
Laptop
AC adaptor
DVD
Docking station
Figure 7.2. Example of an item taxonomy.
milk is the parent of skim milk because there is a directed edge from the ˆ is called an ancestor of X (and X a node milk to the node skim milk. X ˆ ˆ to node X in the directed descendent of X) if there is a path from node X acyclic graph. In the diagram shown in Figure 7.2, food is an ancestor of skim milk and AC adaptor is a descendent of electronics. The main advantages of incorporating concept hierarchies into association analysis are as follows: 1. Items at the lower levels of a hierarchy may not have enough support to appear in any frequent itemsets. For example, although the sale of AC adaptors and docking stations may be low, the sale of laptop accessories, which is their parent node in the concept hierarchy, may be high. Unless the concept hierarchy is used, there is a potential to miss interesting patterns involving the laptop accessories. 2. Rules found at the lower levels of a concept hierarchy tend to be overly speciﬁc and may not be as interesting as rules at the higher levels. For example, staple items such as milk and bread tend to produce many lowlevel rules such as skim milk −→ wheat bread, 2% milk −→ wheat bread, and skim milk −→ white bread. Using a concept hierarchy, they can be summarized into a single rule, milk −→ bread. Considering only items residing at the top level of their hierarchies may not be good enough because such rules may not be of any practical use. For example, although the rule electronics −→ food may satisfy the support and
427
Chapter 7
Association Analysis: Advanced Concepts
conﬁdence thresholds, it is not informative because the combination of electronics and food items that are frequently purchased by customers are unknown. If milk and batteries are the only items sold together frequently, then the pattern {food, electronics} may have overgeneralized the situation. Standard association analysis can be extended to incorporate concept hierarchies in the following way. Each transaction t is initially replaced with its extended transaction t , which contains all the items in t along with their corresponding ancestors. For example, the transaction {DVD, wheat bread} can be extended to {DVD, wheat bread, home electronics, electronics, bread, food}, where home electronics and electronics are the ancestors of DVD, while bread and food are the ancestors of wheat bread. With this approach, existing algorithms such as Apriori can be applied to the extended database to ﬁnd rules that span diﬀerent levels of the concept hierarchy. This approach has several obvious limitations: 1. Items residing at the higher levels tend to have higher support counts than those residing at the lower levels of a concept hierarchy. As a result, if the support threshold is set too high, then only patterns involving the highlevel items are extracted. On the other hand, if the threshold is set too low, then the algorithm generates far too many patterns (most of which may be spurious) and becomes computationally ineﬃcient. 2. Introduction of a concept hierarchy tends to increase the computation time of association analysis algorithms because of the larger number of items and wider transactions. The number of candidate patterns and frequent patterns generated by these algorithms may also grow exponentially with wider transactions. 3. Introduction of a concept hierarchy may produce redundant rules. A ˆ −→ Yˆ , rule X −→ Y is redundant if there exists a more general rule X ˆ is an ancestor of X, Yˆ is an ancestor of Y , and both rules where X have very similar conﬁdence. For example, suppose {bread} −→ {milk}, {white bread} −→ {2% milk}, {wheat bread} −→ {2% milk}, {white bread} −→ {skim milk}, and {wheat bread} −→ {skim milk} have very similar conﬁdence. The rules involving items from the lower level of the hierarchy are considered redundant because they can be summarized by a rule involving the ancestor items. An itemset such as {skim milk, milk, food} is also redundant because food and milk are ancestors of skim milk. Fortunately, it is easy to eliminate such redundant itemsets during frequent itemset generation, given the knowledge of the hierarchy.
428
7.4
Sequential Patterns
10
15
Timeline Sequence Database: Object A A A B B B B C
Timestamp 10 20 23 11 17 21 28 14
Events 2, 3, 5 6, 1 1 4, 5, 6 2 7, 8, 1, 2 1, 6 1, 8, 7
20
25
30
35
Sequence for Object A: 6 1
2 3 5
1
Sequence for Object B: 4 5 6
2
7 8 1 2
1 6
Sequence for Object C: 1 7 8
Figure 7.3. Example of a sequence database.
7.4
Sequential Patterns
Market basket data often contains temporal information about when an item was purchased by customers. Such information can be used to piece together the sequence of transactions made by a customer over a certain period of time. Similarly, eventbased data collected from scientiﬁc experiments or the monitoring of physical systems such as telecommunications networks, computer networks, and wireless sensor networks, have an inherent sequential nature to them. This means that an ordinal relation, usually based on temporal or spatial precedence, exists among events occurring in such data. However, the concepts of association patterns discussed so far emphasize only cooccurrence relationships and disregard the sequential information of the data. The latter information may be valuable for identifying recurring features of a dynamic system or predicting future occurrences of certain events. This section presents the basic concept of sequential patterns and the algorithms developed to discover them.
7.4.1
Problem Formulation
The input to the problem of discovering sequential patterns is a sequence data set, which is shown on the lefthand side of Figure 7.3. Each row records the occurrences of events associated with a particular object at a given time. For example, the ﬁrst row contains the set of events occurring at timestamp t = 10
429
Chapter 7
Association Analysis: Advanced Concepts
for object A. By sorting all the events associated with object A in increasing order of their timestamps, a sequence for object A is obtained, as shown on the righthand side of Figure 7.3. Generally speaking, a sequence is an ordered list of elements. A sequence can be denoted as s = e1 e2 e3 . . . en , where each element ej is a collection of one or more events, i.e., ej = {i1 , i2 , . . . , ik }. The following is a list of examples of sequences: • Sequence of Web pages viewed by a Web site visitor: {Homepage} {Electronics} {Cameras and Camcorders} {Digital Cameras} {Shopping Cart} {Order Conﬁrmation} {Return to Shopping} • Sequence of events leading to the nuclear accident at ThreeMile Island: {clogged resin} {outlet valve closure} {loss of feedwater} {condenser polisher outlet valve shut} {booster pumps trip} {main waterpump trips} {main turbine trips} {reactor pressure increases} • Sequence of classes taken by a computer science major: {Algorithms and Data Structures, Introduction to Operating Systems} {Database Systems, Computer Architecture} {Computer Networks, Software Engineering} {Computer Graphics, Parallel Programming} A sequence can be characterized by its length and the number of occurring events. The length of a sequence corresponds to the number of elements present in the sequence, while a ksequence is a sequence that contains k events. The Web sequence in the previous example contains 7 elements and 7 events; the event sequence at ThreeMile Island contains 8 elements and 8 events; and the class sequence contains 4 elements and 8 events. Figure 7.4 provides examples of sequences, elements, and events deﬁned for a variety of application domains. Except for the last row, the ordinal attribute associated with each of the ﬁrst three domains corresponds to calendar time. For the last row, the ordinal attribute corresponds to the location of the bases (A, C, G, T) in the gene sequence. Although the discussion on sequential patterns is primarily focused on temporal events, it can be extended to the case where the events have spatial ordering. Subsequences A sequence t is a subsequence of another sequence s if each ordered element in t is a subset of an ordered element in s. Formally, the sequence t = t1 t2 . . . tm
430
7.4 Sequence Database
Sequence
Sequential Patterns
Element (Transaction)
Event (Item)
Customer
Purchase history of a given customer
A set of items bought by a customer at time t
Books, diary products, CDs, etc
Web Data
Browsing activity of a particular Web visitor
The collection of files viewed by a Web visitor after a single mouse click
Home page, index page, contact info, etc
Event data
History of events generated by a given sensor
Events triggered by a sensor at time t
Types of alarms generated by sensors
Genome sequences
DNA sequence of a particular species
An element of the DNA sequence
Bases A,T,G,C
Element (Transaction)
Event (Item) E1 E2
E1 E3
E2
E2
E3 E4
Sequence Ordinal Attribute
Figure 7.4. Examples of elements and events in sequence data sets.
is a subsequence of s = s1 s2 . . . sn if there exist integers 1 ≤ j1 < j2 < · · · < jm ≤ n such that t1 ⊆ sj1 , t2 ⊆ sj2 , . . . , tm ⊆ sjm . If t is a subsequence of s, then we say that t is contained in s. The following table gives examples illustrating the idea of subsequences for various sequences. Sequence, s
7.4.2
Sequence, t < {2} {3,6} {8} > < {2} {8} > < {1} {2} > < {2} {4} >
Is t a subsequence of s? Yes Yes No Yes
Sequential Pattern Discovery
Let D be a data set that contains one or more data sequences. The term data sequence refers to an ordered list of events associated with a single data object. For example, the data set shown in Figure 7.3 contains three data sequences, one for each object A, B, and C. The support of a sequence s is the fraction of all data sequences that contain s. If the support for s is greater than or equal to a userspeciﬁed
431
Chapter 7 Object A A A B B C C C D D D E E
Association Analysis: Advanced Concepts
Timestamp 1 2 3 1 2 1 2 3 1 2 3 1 2
Events 1, 2, 4 2, 3 5 1, 2 2, 3, 4 1, 2 2, 3, 4 2, 4, 5 2 3, 4 4, 5 1, 3 2, 4, 5
Minsup = 50% Examples of Sequential Patterns:
s=60% s=60% s=80% s=80% s=80% s=60% s=60% s=60% s=60%
Figure 7.5. Sequential patterns derived from a data set that contains five data sequences.
threshold minsup, then s is declared to be a sequential pattern (or frequent sequence). Deﬁnition 7.1 (Sequential Pattern Discovery). Given a sequence data set D and a userspeciﬁed minimum support threshold minsup, the task of sequential pattern discovery is to ﬁnd all sequences with support ≥ minsup. Figure 7.5 illustrates an example of a data set that contains ﬁve data sequences. The support for the sequence < {1}{2} > is equal to 80% because it occurs in four of the ﬁve data sequences (every object except for D). Assuming that the minimum support threshold is 50%, any sequence that appears in at least three data sequences is considered to be a sequential pattern. Examples of sequential patterns extracted from the given data set include , , , , etc. Sequential pattern discovery is a computationally challenging task because there are exponentially many sequences contained in a given data sequence. For example, the data sequence contains sequences such as , , , etc. It can be easily shown that the total number of ksequences present in a data sequence with n events is nk . A data sequence with nine events therefore contains 9 9 9 + + ... + = 29 − 1 = 511 1 2 9 distinct sequences.
432
7.4
Sequential Patterns
A bruteforce approach for generating sequential patterns is to enumerate all possible sequences and count their respective supports. Given a collection of n events, candidate 1sequences are generated ﬁrst, followed by candidate 2sequences, candidate 3sequences, and so on: 1sequences: 2sequences: 3sequences:
< i1 >, < i2 >, . . ., < in > < {i1 , i2 } >, < {i1 , i3 } >, . . ., < {in−1 , in } >, < {i1 }{i1 } >, < {i1 }{i2 } >,. . ., < {in−1 }{in } > < {i1 , i2 , i3 } >, < {i1 , i2 , i4 } >, . . ., < {i1 , i2 }{i1 } >, . . ., < {i1 }{i1 , i2 } >, . . ., < {i1 }{i1 }{i1 } >, . . ., < {in }{in }{in } >
Notice that the number of candidate sequences is substantially larger than the number of candidate itemsets. There are two reasons for the additional number of candidates: 1. An item can appear at most once in an itemset, but an event can appear more than once in a sequence. Given a pair of items, i1 and i2 , only one candidate 2itemset, {i1 , i2 }, can be generated. On the other hand, there are many candidate 2sequences, such as < {i1 , i2 } >, < {i1 }{i2 } >, < {i2 }{i1 } >, and < {i1 , i1 } >, that can be generated. 2. Order matters in sequences, but not for itemsets. For example, {1, 2} and {2, 1} refers to the same itemset, whereas < {i1 }{i2 } > and < {i2 }{i1 } > correspond to diﬀerent sequences, and thus must be generated separately. The Apriori principle holds for sequential data because any data sequence that contains a particular ksequence must also contain all of its (k − 1)subsequences. An Apriori like algorithm can be developed to extract sequential patterns from a sequence data set. The basic structure of the algorithm is shown in Algorithm 7.1. Notice that the structure of the algorithm is almost identical to Algorithm 6.1 presented in the previous chapter. The algorithm would iteratively generate new candidate ksequences, prune candidates whose (k − 1)sequences are infrequent, and then count the supports of the remaining candidates to identify the sequential patterns. The detailed aspects of these steps are given next. Candidate Generation A pair of frequent (k − 1)sequences are merged to produce a candidate ksequence. To avoid generating duplicate candidates, recall that the traditional Apriori algorithm merges a pair of frequent kitemsets only if their ﬁrst k − 1 items are identical. A similar approach can be used
433
Chapter 7
Association Analysis: Advanced Concepts
Algorithm 7.1 Apriori like algorithm for sequential pattern discovery. 1: 2: 3: 4: 5: 6: 7: 8: 9: 10: 11: 12: 13: 14:
k = 1. ≥ minsup}. {Find all frequent 1subsequences.} Fk = { i  i ∈ I ∧ σ({i}) N repeat k = k + 1. Ck = apriorigen(Fk−1 ). {Generate candidate ksubsequences.} for each data sequence t ∈ T do Ct = subsequence(Ck , t). {Identify all candidates contained in t.} for each candidate ksubsequence c ∈ Ct do σ(c) = σ(c) + 1. {Increment the support count.} end for end for ≥ minsup}. {Extract the frequent ksubsequences.} Fk = { c  c ∈ Ck ∧ σ(c) N until Fk = ∅ Answer = Fk .
for sequences. The criteria for merging sequences are stated in the form of the following procedure.
Sequence Merging Procedure A sequence s(1) is merged with another sequence s(2) only if the subsequence obtained by dropping the ﬁrst event in s(1) is identical to the subsequence obtained by dropping the last event in s(2) . The resulting candidate is the sequence s(1) , concatenated with the last event from s(2) . The last event from s(2) can either be merged into the same element as the last event in s(1) or diﬀerent elements depending on the following conditions: 1. If the last two events in s(2) belong to the same element, then the last event in s(2) is part of the last element in s(1) in the merged sequence. 2. If the last two events in s(2) belong to diﬀerent elements, then the last event in s(2) becomes a separate element appended to the end of s(1) in the merged sequence.
Figure 7.6 illustrates examples of candidate 4sequences obtained by merging pairs of frequent 3sequences. The ﬁrst candidate {1}{2}{3}{4} is obtained by merging (1)(2)(3) with (2)(3)(4). Since events 3 and 4 belong to diﬀerent elements of the second sequence, they also belong to separate elements in the merged sequence. On the other hand, merging {1}{5}{3} with {5}{3, 4} produces the candidate 4sequence {1}{5}{3, 4}. In this case,
434
7.4
Sequential Patterns
Frequent 3sequences < (1) (2) (3) > < (1) (2 5) > < (1) (5) (3) > < (2) (3) (4) > < (2 5) (3) > < (3) (4) (5) > < (5) (3 4) >
Candidate Generation < (1) (2) (3) (4) > < (1) (2 5) (3) > < (1) (5) (3 4) > < (2) (3) (4) (5) > < (2 5) (3 4) >
Candidate Pruning < (1) (2 5) (3) >
Figure 7.6. Example of the candidate generation and pruning steps of a sequential pattern mining algorithm.
since events 3 and 4 belong to the same element of the second sequence, they are combined into the same element in the merged sequence. Finally, the sequences {1}{2}{3} and {1}{2, 5} do not have to be merged because removing the ﬁrst event from the ﬁrst sequence does not give the same subsequence as removing the last event from the second sequence. Although {1}{2, 5}{3} is a viable candidate, it is generated by merging a diﬀerent pair of sequences, {1}{2, 5} and {2, 5}{3}. This example shows that the sequence merging procedure is complete; i.e., it will not miss any viable candidate, while at the same time, it avoids generating duplicate candidate sequences. Candidate Pruning A candidate ksequence is pruned if at least one of its (k −1)sequences is infrequent. For example, suppose {1}{2}{3}{4} is a candidate 4sequence. We need to check whether {1}{2}{4} and {1}{3}{4} are frequent 3sequences. Since both are infrequent, the candidate {1}{2}{3}{4} can be eliminated. Readers should be able to verify that the only candidate 4sequence that survives the candidate pruning step in Figure 7.6 is {1}{2 5}{3}. Support Counting During support counting, the algorithm will enumerate all candidate ksequences belonging to a particular data sequence. The support of these candidates will be incremented. After counting their supports, the algorithm may identify the frequent ksequences and may discard all candidates whose support counts are less than the minsup threshold.
435
Chapter 7
Association Analysis: Advanced Concepts u(sj+1)  I(sj) mingap
Sequence:
1 3 2
1 2 34
window size ws 3 5
2 4 5
u(sn)  I(s1) < {3} {6} > < {1,3} {6} >
Does s support t? Yes Yes No
In general, the longer the maxspan, the more likely it is to detect a pattern in a data sequence. However, a longer maxspan can also capture spurious patterns because it increases the chance for two unrelated events to be temporally related. In addition, the pattern may involve events that are already obsolete. The maxspan constraint aﬀects the support counting step of sequential pattern discovery algorithms. As shown in the preceding examples, some data sequences no longer support a candidate pattern when the maxspan constraint is imposed. If we simply apply Algorithm 7.1, the support counts for some patterns may be overestimated. To avoid this problem, the algorithm must be modiﬁed to ignore cases where the interval between the ﬁrst and last occurrences of events in a given pattern is greater than maxspan. The mingap and maxgap Constraints Timing constraints can also be speciﬁed to restrict the time diﬀerence between two consecutive elements of a sequence. If the maximum time diﬀerence (maxgap) is one week, then events in one element must occur within a week’s time of the events occurring in the previous element. If the minimum time difference (mingap) is zero, then events in one element must occur immediately after the events occurring in the previous element. The following table shows examples of patterns that pass or fail the maxgap and mingap constraints, assuming that maxgap = 3 and mingap = 1. Data Sequence, s
Sequential Pattern, t < {3} {6} > < {6} {8} > < {1,3} {6} > < {1} {3} {8} >
maxgap Pass Pass Fail Fail
mingap Pass Fail Pass Fail
437
Chapter 7
Association Analysis: Advanced Concepts
As with maxspan, these constraints will aﬀect the support counting step of sequential pattern discovery algorithms because some data sequences no longer support a candidate pattern when mingap and maxgap constraints are present. These algorithms must be modiﬁed to ensure that the timing constraints are not violated when counting the support of a pattern. Otherwise, some infrequent sequences may mistakenly be declared as frequent patterns. A side eﬀect of using the maxgap constraint is that the Apriori principle might be violated. To illustrate this, consider the data set shown in Figure 7.5. Without mingap or maxgap constraints, the support for {2}{5} and {2}{3}{5} are both equal to 60%. However, if mingap = 0 and maxgap = 1, then the support for {2}{5} reduces to 40%, while the support for {2}{3}{5} is still 60%. In other words, support has increased when the number of events in a sequence increases—which contradicts the Apriori principle. The violation occurs because the object D does not support the pattern {2}{5} since the time gap between events 2 and 5 is greater than maxgap. This problem can be avoided by using the concept of a contiguous subsequence. Deﬁnition 7.2 (Contiguous Subsequence). A sequence s is a contiguous subsequence of w = e1 e2 . . . ek if any one of the following conditions hold: 1. s is obtained from w after deleting an event from either e1 or ek , 2. s is obtained from w after deleting an event from any element ei ∈ w that contains at least two events, or 3. s is a contiguous subsequence of t and t is a contiguous subsequence of w. The following examples illustrate the concept of a contiguous subsequence: Data Sequence, s
Sequential Pattern, t
< < < <
> > > >
Is t a contiguous subsequence of s? Yes Yes Yes No No
Using the concept of contiguous subsequences, the Apriori principle can be modiﬁed to handle maxgap constraints in the following way. Deﬁnition 7.3 (Modiﬁed Apriori Principle). If a ksequence is frequent, then all of its contiguous k − 1subsequences must also be frequent.
438
7.4
Sequential Patterns
The modiﬁed Apriori principle can be applied to the sequential pattern discovery algorithm with minor modiﬁcations. During candidate pruning, not all ksequences need to be veriﬁed since some of them may violate the maxgap constraint. For example, if maxgap = 1, it is not necessary to check whether the subsequence {1}{2, 3}{5} of the candidate {1}{2, 3}{4}{5} is frequent since the time diﬀerence between elements {2, 3} and {5} is greater than one time unit. Instead, only the contiguous subsequences of {1}{2, 3}{4}{5} need to be examined. These subsequences include {1}{2, 3}{4}, {2, 3}{4}{5}, {1}{2}{4}{5}, and {1}{3}{4}{5}. The Window Size Constraint Finally, events within an element sj do not have to occur at the same time. A window size threshold (ws) can be deﬁned to specify the maximum allowed time diﬀerence between the latest and earliest occurrences of events in any element of a sequential pattern. A window size of 0 means all events in the same element of a pattern must occur simultaneously. The following example uses ws = 2 to determine whether a data sequence supports a given sequence (assuming mingap = 0, maxgap = 3, and maxspan = ∞). Data Sequence, s
Sequential Pattern, t < {3,4} {5} > < {4,6} {8} > < {3, 4, 6} {8} > < {1,3,4} {6,7,8} >
Does s support t? Yes Yes No No
In the last example, although the pattern {1,3,4} {6,7,8} satisﬁes the window size constraint, it violates the maxgap constraint because the maximum time diﬀerence between events in the two elements is 5 units. The window size constraint also aﬀects the support counting step of sequential pattern discovery algorithms. If Algorithm 7.1 is applied without imposing the window size constraint, the support counts for some of the candidate patterns might be underestimated, and thus some interesting patterns may be lost.
7.4.4
Alternative Counting Schemes
There are several methods available for counting the support of a candidate ksequence from a database of sequences. For illustrative purposes, consider the problem of counting the support for sequence {p}{q}, as shown in Figure 7.8. Assume that ws = 0, mingap = 0, maxgap = 1, and maxspan = 2.
439
Chapter 7
Association Analysis: Advanced Concepts
Objectʼs Timeline p p p q q 1
2
3
4
p q 5
p q 6
Sequence: (p) (q) q
(Method, Count) 7 COBJ
1
CWIN
6
CMINWIN
4
CDIST_O
8
CDIST
5
Figure 7.8. Comparing different support counting methods.
• COBJ: One occurrence per object. This method looks for at least one occurrence of a given sequence in an object’s timeline. In Figure 7.8, even though the sequence (p)(q) appears several times in the object’s timeline, it is counted only once— with p occurring at t = 1 and q occuring at t = 3. • CWIN: One occurrence per sliding window. In this approach, a sliding time window of ﬁxed length (maxspan) is moved across an object’s timeline, one unit at a time. The support count is incremented each time the sequence is encountered in the sliding window. In Figure 7.8, the sequence {p}{q} is observed six times using this method. • CMINWIN: Number of minimal windows of occurrence. A minimal window of occurrence is the smallest window in which the sequence occurs given the timing constraints. In other words, a minimal
440
7.4
Sequential Patterns
window is the time interval such that the sequence occurs in that time interval, but it does not occur in any of the proper subintervals of it. This deﬁnition can be considered as a restrictive version of CWIN, because its eﬀect is to shrink and collapse some of the windows that are counted by CWIN. For example, sequence {p}{q} has four minimal window occurrences: (1) the pair (p: t = 2, q: t = 3), (2) the pair (p: t = 3, q: t = 4), (3) the pair (p: t = 5, q: t = 6), and (4) the pair (p: t = 6, q: t = 7). The occurrence of event p at t = 1 and event q at t = 3 is not a minimal window occurrence because it contains a smaller window with (p: t = 2, q: t = 3), which is indeed a minimal window of occurrence. • CDIST O: Distinct occurrences with possibility of eventtimestamp overlap. A distinct occurrence of a sequence is deﬁned to be the set of eventtimestamp pairs such that there has to be at least one new eventtimestamp pair that is diﬀerent from a previously counted occurrence. Counting all such distinct occurrences results in the CDIST O method. If the occurrence time of events p and q is denoted as a tuple (t(p), t(q)), then this method yields eight distinct occurrences of sequence {p}{q} at times (1,3), (2,3), (2,4), (3,4), (3,5), (5,6), (5,7), and (6,7). • CDIST: Distinct occurrences with no eventtimestamp overlap allowed. In CDIST O above, two occurrences of a sequence were allowed to have overlapping eventtimestamp pairs, e.g., the overlap between (1,3) and (2,3). In the CDIST method, no overlap is allowed. Eﬀectively, when an eventtimestamp pair is considered for counting, it is marked as used and is never used again for subsequent counting of the same sequence. As an example, there are ﬁve distinct, nonoverlapping occurrences of the sequence {p}{q} in the diagram shown in Figure 7.8. These occurrences happen at times (1,3), (2,4), (3,5), (5,6), and (6,7). Observe that these occurrences are subsets of the occurrences observed in CDIST O. One ﬁnal point regarding the counting methods is the need to determine the baseline for computing the support measure. For frequent itemset mining, the baseline is given by the total number of transactions. For sequential pattern mining, the baseline depends on the counting method used. For the COBJ method, the total number of objects in the input data can be used as the baseline. For the CWIN and CMINWIN methods, the baseline is given by the sum of the number of time windows possible in all objects. For methods such as CDIST and CDIST O, the baseline is given by the sum of the number of distinct timestamps present in the input data of each object.
441
Chapter 7
7.5
Association Analysis: Advanced Concepts
Subgraph Patterns
This section describes the application of association analysis methods to more complex entities beyond itemsets and sequences. Examples include chemical compounds, 3D protein structures, network topologies, and tree structured XML documents. These entities can be modeled using a graph representation, as shown in Table 7.7. Table 7.7. Graph representation of entities in various application domains. Application Web mining Computational chemistry Network computing
Graphs Web browsing patterns Structure of chemical compounds Computer networks
Semantic Web
Collection of XML documents Protein structures
Bioinformatics
Vertices Web pages Atoms or ions Computers and servers XML elements Amino acids
Edges Hyperlink between pages Bond between atoms or ions Interconnection between machines Parentchild relationship between elements Contact residue
A useful data mining task to perform on this type of data is to derive a set of common substructures among the collection of graphs. Such a task is known as frequent subgraph mining. A potential application of frequent subgraph mining can be seen in the context of computational chemistry. Each year, new chemical compounds are designed for the development of pharmaceutical drugs, pesticides, fertilizers, etc. Although the structure of a compound is known to play a major role in determining its chemical properties, it is difﬁcult to establish their exact relationship. Frequent subgraph mining can aid this undertaking by identifying the substructures commonly associated with certain properties of known compounds. Such information can help scientists to develop new chemical compounds that have certain desired properties. This section presents a methodology for applying association analysis to graphbased data. The section begins with a review of some of the basic graphrelated concepts and deﬁnitions. The frequent subgraph mining problem is then introduced, followed by a description of how the traditional Apriori algorithm can be extended to discover such patterns.
442
7.5
7.5.1
Subgraph Patterns
Graphs and Subgraphs
A graph is a data structure that can be used to represent the relationships among a set of entities. Mathematically, a graph is composed of a vertex set V and a set of edges E connecting between pairs of vertices. Each edge is denoted by a vertex pair (vi , vj ), where vi , vj ∈ V . A label l(vi ) can be assigned to each vertex vi representing the name of an entity. Similarly each edge (vi , vj ) can also be associated with a label l(vi , vj ) describing the relationship between a pair of entities. Table 7.7 shows the vertices and edges associated with diﬀerent types of graphs. For example, in a Web graph, the vertices correspond to Web pages and the edges represent the hyperlinks between Web pages. Deﬁnition 7.4 (Subgraph). A graph G = (V , E ) is a subgraph of another graph G = (V, E) if its vertex set V is a subset of V and its edge set E is a subset of E. The subgraph relationship is denoted as G ⊆S G. Figure 7.9 shows a graph that contains 6 vertices and 11 edges along with one of its possible subgraphs. The subgraph, which is shown in Figure 7.9(b), contains only 4 of the 6 vertices and 4 of the 11 edges in the original graph. a
a
q
p p
b
p a
s
s
a
r t
p
t
t
r c
c p
q
c p
b
b
(a) Labeled graph.
(b) Subgraph.
Figure 7.9. Example of a subgraph.
Deﬁnition 7.5 (Support). Given a collection of graphs G, the support for a subgraph g is deﬁned as the fraction of all graphs that contain g as its subgraph, i.e.: {Gi g ⊆S Gi , Gi ∈ G} . (7.2) s(g) = G
443
Chapter 7 e
1
a 1 b
Association Analysis: Advanced Concepts
1
1 d
1
d
1
1 c
1 e
1
c
e
support = 80%
Subgraph g2
G2
1
1
b 1 a
e 1 d
1
1
a
b
G1 a
Subgraph g1
1
a
G3
d
1
a
1 e
1 e
1 1
d
support = 60% c
G4 d
b 1 a
1 1
1
e
G5
Subgraph g3
c
Graph Data Set
1
a 1
d 1 e
support = 40%
Figure 7.10. Computing the support of a subgraph from a set of graphs.
Example 7.2. Consider the ﬁve graphs, G1 through G5 , shown in Figure 7.10. The graph g1 shown on the top righthand diagram is a subgraph of G1 , G3 , G4 , and G5 . Therefore, s(g1 ) = 4/5 = 80%. Similarly, we can show that s(g2 ) = 60% because g2 is a subgraph of G1 , G2 , and G3 , while s(g3 ) = 40% because g3 is a subgraph of G1 and G3 .
7.5.2
Frequent Subgraph Mining
This section presents a formal deﬁnition of the frequent subgraph mining problem and illustrates the complexity of this task. Deﬁnition 7.6 (Frequent Subgraph Mining). Given a set of graphs G and a support threshold, minsup, the goal of frequent subgraph mining is to ﬁnd all subgraphs g such that s(g) ≥ minsup. While this formulation is generally applicable to any type of graph, the discussion presented in this chapter focuses primarily on undirected, connected graphs. The deﬁnitions of these graphs are given below: 1. A graph is connected if there exists a path between every pair of vertices in the graph, in which a path is a sequence of vertices < v1 v2 . . . vk >
444
7.5
Subgraph Patterns
such that there is an edge connecting between every pair of adjacent vertices (vi , vi+1 ) in the sequence. 2. A graph is undirected if it contains only undirected edges. An edge (vi , vj ) is undirected if it is indistinguishable from (vj , vi ). Methods for handling other types of subgraphs (directed or disconnected) are left as an exercise to the readers (see Exercise 15 on page 482). Mining frequent subgraphs is a computationally expensive task because of the exponential scale of the search space. To illustrate the complexity of this task, consider a data set that contains d entities. In frequent itemset mining, each entity is an item and the size of the search space to be explored is 2d , which is the number of candidate itemsets that can be generated. In frequent subgraph mining, each entity is a vertex and can have up to d − 1 edges to other vertices. Assuming that the vertex labels are unique, the total number of subgraphs is d d × 2i(i−1)/2 , i i=1 d where i is the number of ways to choose i vertices to form a subgraph and 2i(i−1)/2 is the maximum number of edges between vertices. Table 7.8 compares the number of itemsets and subgraphs for diﬀerent values of d. Table 7.8. A comparison between number of itemsets and subgraphs for different dimensionality, d. Number of entities, d Number of itemsets Number of subgraphs
1 2 2
2 4 5
3 8 18
4 16 113
5 32 1,450
6 64 40,069
7 128 2,350,602
8 256 28,619,2513
The number of candidate subgraphs is actually much smaller because the numbers given in Table 7.8 include subgraphs that are disconnected. Disconnected subgraphs are usually ignored because they are not as interesting as connected subgraphs. A bruteforce method for doing this is to generate all connected subgraphs as candidates and count their respective supports. For example, consider the graphs shown in Figure 7.11(a). Assuming that the vertex labels are chosen from the set {a, b} and the edge labels are chosen from the set {p, q}, the list of connected subgraphs with one vertex up to three vertices is shown in Figure 7.11(b). The number of candidate subgraphs is considerably larger than the
445
Chapter 7 a
Association Analysis: Advanced Concepts
a
p p
b
p
p
q
b
p a
a
G2
b
q
p
a
a G1
b
p
q
p
p
b
a
q
a
p
p
b q
b
a
q
q
b G3
G4
(a) Example of a graph data set.
a
k=1 k=2
a a b
k=3
a
b p p p p
a
q
a a
b b a
a
q
b p
a
b
q
b a
p
q
a
a
p
p a
p
a
b
a p
q
b q
a
a
p
...
q
b
...
a
q
b
b q
q b
(b) List of connected subgraphs.
Figure 7.11. Bruteforce method for mining frequent subgraphs.
number of candidate itemsets in traditional association rule mining for the following reasons: 1. An item can appear at most once in an itemset, whereas a vertex label can appear more than once in a graph. 2. The same pair of vertex labels can have multiple choices of edge labels. Given the large number of candidate subgraphs, a bruteforce method may break down even for moderately sized graphs.
446
7.5 a
a
p p
b
p a
p
G1 G2 G3 G4
a G2 (a,b,p) (a,b,q) 1 0 1 0 0 0 0 0
p
b
b
(a,b,r) 0 0 1 0
q
b
q
p a
a
b G3
(b,c,p) (b,c,q) 0 0 0 0 1 0 0 0
b
q
p
p
q
a G1
a
p
p
p b
q
b
a
q
Subgraph Patterns
q G4
(b,c,r) 1 0 0 0
... ... ... ... ...
(d,e,r) 0 0 0 0
Figure 7.12. Mapping a collection of graph structures into market basket transactions.
7.5.3
Apriori like Method
This section examines how an Apriori like algorithm can be developed for ﬁnding frequent subgraphs. Data Transformation One possible approach is to transform each graph into a transactionlike format so that existing algorithms such as Apriori can be applied. Figure 7.12 illustrates how to transform a collection of graphs into its equivalent market basket representation. In this representation, each combination of edge label l(e) with its corresponding vertex labels, (l(vi ), l(vj )), is mapped into an “item.” The width of the “transaction” is given by the number of edges in the graph. Despite its simplicity, this approach works only if every edge in a graph has a unique combination of vertex and edge labels. Otherwise, such graphs cannot be accurately modeled using this representation. General Structure of the Frequent Subgraph Mining Algorithm An Apriori like algorithm for mining frequent subgraphs consists of the following steps: 1. Candidate generation, which is the process of merging pairs of frequent (k − 1)subgraphs to obtain a candidate ksubgraph.
447
Chapter 7
Association Analysis: Advanced Concepts
2. Candidate pruning, which is the process of discarding all candidate ksubgraphs that contain infrequent (k − 1)subgraphs. 3. Support counting, which is the process of counting the number of graphs in G that contain each candidate. 4. Candidate elimination, which discards all candidate subgraphs whose support counts are less than minsup. The speciﬁc details of these steps are discussed in the remainder of this section.
7.5.4
Candidate Generation
During candidate generation, a pair of frequent (k − 1)subgraphs are merged to form a candidate ksubgraph. The ﬁrst question is how to deﬁne k, the size of a subgraph. In the example shown in Figure 7.11, k refers to the number of vertices in the graph. This approach of iteratively expanding a subgraph by adding an extra vertex is known as vertex growing. Alternatively, k may refer to the number of edges in the graph. This approach of adding an extra edge to the existing subgraphs is known as edge growing. To avoid generating duplicate candidates, we may impose an additional condition for merging, that the two (k − 1)subgraphs must share a common (k−2)subgraph. The common (k−2)subgraph is known as their core. Below, we brieﬂy describe the candidate generation procedure for both vertexgrowing and edgegrowing strategies. Candidate Generation via Vertex Growing Vertex growing is the process of generating a new candidate by adding a new vertex into an existing frequent subgraph. Before describing this approach, let us ﬁrst consider the adjacency matrix representation of a graph. Each entry M (i, j) in the matrix contains either the label of the edge connecting between the vertices vi and vj , or zero, if there is no edge between them. The vertexgrowing approach can be viewed as the process of generating a k × k adjacency matrix by combining a pair of (k − 1) × (k − 1) adjacency matrices, as illustrated in Figure 7.13. G1 and G2 are two graphs whose adjacency matrices are given by M (G1) and M (G2), respectively. The core for the graphs is indicated by dashed lines in the diagram. The procedure for generating candidate subgraphs via vertex growing is presented next.
448
7.5 a e
p
p
p
a G1
G2
r
d
a G3 = merge (G1, G2)
0
p
p
q
0
p
p
0
p
0
r
0
p
0
r
0
p
r
0
0
p
r
0
r
q
0
0
0
0
0
r
0
MG2 =
?
r
r a
p
a
d r
q e
p
p
a
q
MG1 =
a
a
q
b
Subgraph Patterns
0
p
p
q
0
p
0
r
0
0
MG3 = p
r
0
0
r
q
0
0
0
?
0
0
r
?
0
Figure 7.13. Vertexgrowing strategy.
Subgraph Merging Procedure via Vertex Growing An adjacency matrix M (1) is merged with another matrix M (2) if the submatrices obtained by removing the last row and last column of M (1) and M (2) are identical to each other. The resulting matrix is the matrix M (1) , appended with the last row and last column of matrix M (2) . The remaining entries of the new matrix are either zero or replaced by all valid edge labels connecting the pair of vertices.
The resulting graph contains one or two edges more than the original graphs. In Figure 7.13, both G1 and G2 contain four vertices and four edges. After merging, the resulting graph G3 has ﬁve vertices. The number of edges in G3 depends on whether the vertices d and e are connected. If d and e are disconnected, then G3 has ﬁve edges and the corresponding matrix entry for (d, e) is zero. Otherwise, G3 has six edges and the matrix entry for (d, e) corresponds to the label for the newly created edge. Since the edge label is unknown, we need to consider all possible edge labels for (d, e), thus increasing the number of candidate subgraphs substantially. Candidate Generation via Edge Growing Edge growing inserts a new edge to an existing frequent subgraph during candidate generation. Unlike vertex growing, the resulting subgraph does not
449
Chapter 7
Association Analysis: Advanced Concepts a
q e
p a e
p
p
e r
G3 = merge (G1, G2)
r
a
a
G1
G2
e
a
a q
r
r p
p
b
a
a
q
p
a
q e
p
p
a r
r a
G4 = merge (G1, G2)
Figure 7.14. Edgegrowing strategy.
necessarily increase the number of vertices in the original graphs. Figure 7.14 shows two possible candidate subgraphs obtained by merging G1 and G2 via the edgegrowing strategy. The ﬁrst candidate subgraph, G3, has one extra vertex, while the second candidate subgraph, G4, has the same number of vertices as the original graphs. The core for the graphs is indicated by dashed lines in the diagram. The procedure for generating candidate subgraphs via edge growing can be summarized as follows. Subgraph Merging Procedure via Edge Growing A frequent subgraph g (1) is merged with another frequent subgraph g (2) only if the subgraph obtained by removing an edge from g (1) is topologically equivalent to the subgraph obtained by removing an edge from g (2) . After merging, the resulting candidate is the subgraph g (1) , appended with the extra edge from g (2) .
The graphs to be merged may contain several vertices that are topologically equivalent to each other. To illustrate the concept of topologically equivalent vertices, consider the graphs shown in Figure 7.15. The graph G1 contains four vertices with identical vertex labels, “a.” If a new edge is at
450
7.5 v1 a
p
p
v1
v2 a
a p
p
p p
Subgraph Patterns
v2 a
v1 p a
p
v2
p
v3
b
p
b
v4 a
p a v3
p
a v4
a v3
p
a v4
b v5
G2
G1
G3
Figure 7.15. Illustration of topologically equivalent vertices.
tached to any one of the four vertices, the resulting graph will look the same. The vertices in G1 are therefore topologically equivalent to each other. The graph G2 has two pairs of topologically equivalent vertices, v1 with v4 and v2 with v3 , even though the vertex and edge labels are identical. It is easy to see that v1 is not topologically equivalent to v2 because the number of edges incident on the vertices is diﬀerent. Therefore, attaching a new edge to v1 results in a diﬀerent graph than attaching the same edge to v2 . Meanwhile, the graph G3 does not have any topologically equivalent vertices. While v1 and v4 have the same vertex labels and number of incident edges, attaching a new edge to v1 results in a diﬀerent graph than attaching the same edge to v4 . The notion of topologically equivalent vertices can help us understand why multiple candidate subgraphs can be generated during edge growing. Consider the (k − 1)subgraphs G1 and G2 shown in Figure 7.16. To simplify the notation, their core, which contains k − 2 common edges between the two graphs, is drawn as a rectangular box. The remaining edge in G1 that is not included in the core is shown as a dangling edge connecting the vertices a and b. Similarly, the remaining edge in G2 that is not part of the core is shown as a dangling edge connecting vertices c and d. Although the cores for G1 and G2 are identical, a and c may or may not be topologically equivalent to each G1
G2 a
Core
b
c
d
Core
Figure 7.16. General approach for merging a pair of subgraphs via edge growing.
451
Chapter 7
Association Analysis: Advanced Concepts
G3 = Merge (G1, G2)
Core
G3 = Merge (G1, G2)
a
b
c
d
Core
(a) a ≠ c and b ≠ d
b
c
d
a
b
Core
d
(b) a = c and b ≠ d
G3 = Merge (G1, G2)
Core
G3 = Merge (G1, G2)
a
G3 = Merge (G1, G2)
a
b
c
d
a
Core
b
c
(c) a ≠ c and b = d G3 = Merge (G1, G2)
Core G3 = Merge (G1, G2) a
Core
a
b
c
d G3 = Merge (G1, G2)
b
d
a
Core
b
c
(d) a = c and b = d
Figure 7.17. Candidate subgraphs generated via edge growing.
other. If a and c are topologically equivalent, we denote them as a = c. For vertices outside the core, we denote them as b = d if their labels are identical. The following rule of thumb can be used to determine the candidate subgraphs obtained during candidate generation: 1. If a = c and b = d, then there is only one possible resulting subgraph, as shown in Figure 7.17(a). 2. If a = c but b = d, then there are two possible resulting subgraphs, as shown in Figure 7.17(b).
452
7.5 a
b
a
b
a
a
b
a
a
a
a
a
a
a
b
a
a
Subgraph Patterns a
b
a
a
b
a
a
b
a
Figure 7.18. Multiplicity of candidates during candidate generation.
3. If a = c but b = d, then there are two possible resulting subgraphs, as shown in Figure 7.17(c). 4. If a = c and b = d, then there are three possible resulting subgraphs, as shown in Figure 7.17(d). Multiple candidate subgraphs can also be generated when there is more than one core associated with the pair of (k −1)subgraphs, as shown in Figure 7.18. The shaded vertices correspond to those vertices whose edges form a core during the merging operation. Each core may lead to a diﬀerent set of candidate subgraphs. In principle, if a pair of frequent (k − 1)subgraphs is merged, there can be at most k−2 cores, each of which is obtained by removing an edge from one of the merged graphs. Although the edgegrowing procedure can produce multiple candidate subgraphs, the number of candidate subgraphs tends to be smaller than those produced by the vertexgrowing strategy.
7.5.5
Candidate Pruning
After the candidate ksubgraphs are generated, the candidates whose (k − 1)subgraphs are infrequent need to be pruned. The pruning step can be performed by successively removing an edge from the candidate ksubgraph and checking whether the corresponding (k − 1)subgraph is connected and frequent. If not, the candidate ksubgraph can be discarded. To check whether the (k − 1)subgraph is frequent, it should be matched against other frequent (k −1)subgraphs. Determining whether two graphs are topologically equivalent (or isomorphic) is known as the graph isomorphism problem. To illustrate the diﬃculty of solving the graph isomorphism problem,
453
Chapter 7
Association Analysis: Advanced Concepts B A
A B
B
A
B A
B
B
A
A
A A
B B
Figure 7.19. Graph isomorphism
consider the two graphs shown in Figure 7.19. Even though both graphs look diﬀerent, they are actually isomorphic to each other because there is a onetoone mapping between vertices in both graphs. Handling Graph Isomorphism A standard approach for handling the graph isomorphism problem is to map each graph into a unique string representation known as its code or canonical label. A canonical label has the property that if two graphs are isomorphic, then their codes must be the same. This property allows us to test for graph isomorphism by comparing the canonical labels of the graphs. The ﬁrst step toward constructing the canonical label of a graph is to ﬁnd an adjacency matrix representation for the graph. Figure 7.20 shows an
a
q e
p p
a r a
M=
0
p
p q
p
0
r
0
p
r
0
0
q
0
0
0
Figure 7.20. Adjacency matrix representation of a graph.
454
7.5
Subgraph Patterns
example of such a matrix for the given graph. In principle, a graph can have more than one adjacency matrix representation because there are multiple ways to order the vertices in the adjacency matrix. In the example shown in Figure 7.20, the ﬁrst row and column correspond to the vertex a that has 3 edges, the second row and column correspond to another vertex a that has 2 edges, and so on. To derive all the adjacency matrix representations for a graph, we need to consider all possible permutations of rows (and their corresponding columns) of the matrix. Mathematically, each permutation corresponds to a multiplication of the initial adjacency matrix with a corresponding permutation matrix, as illustrated in the following example. Example 7.3. Consider the following matrix:
1 2 3 4 5 6 7 8 M = 9 10 11 12 13 14 15 16 The following permutation matrix can be used to exchange the ﬁrst row (and column) with the third row (and column) of M :
P13
0 0 = 1 0
0 1 0 0
1 0 0 0
0 0 , 0 1
where P13 is obtained by swapping the ﬁrst and third row of the identity matrix. To exchange the ﬁrst and third rows (and columns), the permutation matrix is multiplied with M :
T × M × P13 M = P13
11 10 9 12 7 6 5 8 = 3 2 1 4 . 15 14 13 16
Note that multiplying M from the right with P13 exchanges the ﬁrst and third T exchanges the ﬁrst columns of M , while multiplying M from the left with P13 and third rows of M . If all three matrices are multiplied, this will produce a matrix M whose ﬁrst and third rows and columns have been swapped.
455
Chapter 7 A(1)
Association Analysis: Advanced Concepts
A(2)
B (5)
B (6)
B (7)
B (8)
A(3)
A(4)
A(2)
A(1)
B (7)
B (6)
B (5)
B (8)
A(3)
A(4)
A(1) A(2) A(3) A(4) B(5) B(6) B(7) B(8)
A(1) A(2) A(3) A(4) B(5) B(6) B(7) B(8) 0 1 1 1 0 0 0 0 1 0 0 0 0 1 1 0 1 0 0 0 0 1 0 1 0 1 1 0 1 0 0 0 1 0 0 0 0 0 1 1 0 1 0 1 1 0 0 0 0 0 1 1 1 0 0 0 0 0 0 0 0 1 1 1 Code = 1100111000010010010100001011
A(1) A(2) A(3) A(4) B(5) B(6) B(7) B(8)
A(1) A(2) A(3) A(4) B(5) B(6) B(7) B(8) 0 1 1 0 0 0 1 0 1 0 0 0 0 1 0 1 0 0 1 0 1 0 1 0 0 0 0 1 0 1 0 1 1 0 0 1 0 1 0 0 1 0 0 1 0 0 0 1 0 1 0 0 1 0 1 0 0 1 1 0 1 0 0 0 Code = 1011010010100000100110001110
Figure 7.21. String representation of adjacency matrices.
The second step is to determine the string representation for each adjacency matrix. Since the adjacency matrix is symmetric, it is suﬃcient to construct the string representation based on the upper triangular part of the matrix. In the example shown in Figure 7.21, the code is obtained by concatenating the entries of the upper triangular matrix in a columnwise fashion. The ﬁnal step is to compare all the string representations of the graph and choose the one that has the lowest (or highest) lexicographic value. The preceding approach seems expensive because it requires us to examine all possible adjacency matrices of a graph and to compute each of their string representation in order to ﬁnd the canonical label. More speciﬁcally, there are k! permutations that must be considered for every graph that contains k vertices. Some of the methods developed to reduce the complexity of this task include caching the previously computed canonical label (so that we do not have to recompute it again when performing an isomorphism test on the same graph) and reducing the number of permutations needed to determine the canonical label by incorporating additional information such as vertex labels and the degree of a vertex. The latter approach is beyond the scope of this
456
7.6
Infrequent Patterns
book, but interested readers may consult the bibliographic notes at the end of this chapter.
7.5.6
Support Counting
Support counting is also a potentially costly operation because all the candidate subgraphs contained in each graph G ∈ G must be determined. One way to speed up this operation is to maintain a list of graph IDs associated with each frequent (k − 1)subgraph. Whenever a new candidate ksubgraph is generated by merging a pair of frequent (k − 1)subgraphs, their corresponding lists of graph IDs are intersected. Finally, the subgraph isomorphism tests are performed on the graphs in the intersected list to determine whether they contain a particular candidate subgraph.
7.6
Infrequent Patterns
The association analysis formulation described so far is based on the premise that the presence of an item in a transaction is more important than its absence. As a consequence, patterns that are rarely found in a database are often considered to be uninteresting and are eliminated using the support measure. Such patterns are known as infrequent patterns. Deﬁnition 7.7 (Infrequent Pattern). An infrequent pattern is an itemset or a rule whose support is less than the minsup threshold. Although a vast majority of infrequent patterns are uninteresting, some of them might be useful to the analysts, particularly those that correspond to negative correlations in the data. For example, the sale of DVDs and VCRs together is low because any customer who buys a DVD will most likely not buy a VCR, and vice versa. Such negativecorrelated patterns are useful to help identify competing items, which are items that can be substituted for one another. Examples of competing items include tea versus coﬀee, butter versus margarine, regular versus diet soda, and desktop versus laptop computers. Some infrequent patterns may also suggest the occurrence of interesting rare events or exceptional situations in the data. For example, if {Fire = Yes} is frequent but {Fire = Yes, Alarm = On} is infrequent, then the latter is an interesting infrequent pattern because it may indicate faulty alarm systems. To detect such unusual situations, the expected support of a pattern must be determined, so that, if a pattern turns out to have a considerably lower support than expected, it is declared as an interesting infrequent pattern.
457
Chapter 7
Association Analysis: Advanced Concepts
Mining infrequent patterns is a challenging endeavor because there is an enormous number of such patterns that can be derived from a given data set. More speciﬁcally, the key issues in mining infrequent patterns are: (1) how to identify interesting infrequent patterns, and (2) how to eﬃciently discover them in large data sets. To get a diﬀerent perspective on various types of interesting infrequent patterns, two related concepts—negative patterns and negatively correlated patterns—are introduced in Sections 7.6.1 and 7.6.2, respectively. The relationships among these patterns are elucidated in Section 7.6.3. Finally, two classes of techniques developed for mining interesting infrequent patterns are presented in Sections 7.6.5 and 7.6.6.
7.6.1
Negative Patterns
Let I = {i1 , i2 , . . . , id } be a set of items. A negative item, ik , denotes the absence of item ik from a given transaction. For example, coffee is a negative item whose value is 1 if a transaction does not contain coffee. Deﬁnition 7.8 (Negative Itemset). A negative itemset X is an itemset that has the following properties: (1) X = A ∪ B, where A is a set of positive items, B is a set of negative items, B ≥ 1, and (2) s(X) ≥ minsup. Deﬁnition 7.9 (Negative Association Rule). A negative association rule is an association rule that has the following properties: (1) the rule is extracted from a negative itemset, (2) the support of the rule is greater than or equal to minsup, and (3) the conﬁdence of the rule is greater than or equal to minconf. The negative itemsets and negative association rules are collectively known as negative patterns throughout this chapter. An example of a negative association rule is tea −→ coffee, which may suggest that people who drink tea tend to not drink coﬀee.
7.6.2
Negatively Correlated Patterns
Section 6.7.1 on page 371 described how correlation analysis can be used to analyze the relationship between a pair of categorical variables. Measures such as interest factor (Equation 6.5) and the φcoeﬃcient (Equation 6.8) were shown to be useful for discovering itemsets that are positively correlated. This section extends the discussion to negatively correlated patterns. Let X = {x1 , x2 , . . . , xk } denote a kitemset and P (X) denote the probability that a transaction contains X. In association analysis, the probability is often estimated using the itemset support, s(X).
458
7.6
Infrequent Patterns
Deﬁnition 7.10 (Negatively Correlated Itemset). An itemset X is negatively correlated if
s(X)